CompTIA AI+ AI0-001 AI0-001 Questions 751–825 | Page 11/14

751

MCQeasy

An e-commerce company deploys a model to recommend products to users. The recommendation system uses collaborative filtering based on user-item interaction history. After deployment, the model shows decreasing click-through rates (CTR) over time. The data engineer notices that the model was trained on data from the past six months and is retrained daily. However, the trend suggests that user preferences are shifting more rapidly than expected. The engineer suspects that the model is suffering from distribution drift. Which approach should the engineer implement to adapt the model more quickly to changing user behavior?

A.Increase the retraining period to once per week to reduce computational cost

B.Switch to an online learning algorithm that updates the model after each user click

C.Increase the model complexity by adding more features and layers

D.Use only the last week of data for training to focus on recent trends

AnswerB

Online learning continuously adapts to new data, capturing shifts in user preferences promptly.

Why this answer

Option A is correct. Online learning allows the model to update incrementally with each new interaction, adapting quickly to changes. Option B is wrong because batch retraining weekly is slower than daily.

Option C is wrong because using only last week's data may not provide enough data and could be noisy. Option D is wrong because increasing model complexity may cause overfitting and is not a direct solution to drift.

Full explanation →

752

MCQeasy

A security analyst discovers that an attacker has been querying a production LLM API with thousands of carefully crafted prompts and using the responses to build a local copy of the model. Which attack is occurring?

A.Prompt injection

B.Model extraction

C.Data poisoning

D.Membership inference

AnswerB

Model extraction uses queries to approximate the model's behavior and steal its functionality.

Why this answer

Model extraction attacks involve querying a model to reconstruct its functionality or parameters. Membership inference determines if data was in training, data poisoning corrupts training, and prompt injection manipulates prompts.

Full explanation →

753

Multi-Selectmedium

Which TWO techniques are commonly used for feature selection in machine learning? (Choose 2)

Select 2 answers

A.Principal Component Analysis (PCA)

B.SMOTE

C.L1 regularization (Lasso)

D.Dropout

E.Recursive Feature Elimination (RFE)

AnswersC, E

Lasso can zero out coefficients.

Why this answer

L1 regularization (Lasso) is correct because it adds a penalty equal to the absolute value of the magnitude of coefficients, which can shrink some coefficients exactly to zero, effectively performing feature selection by removing irrelevant features from the model. This makes it a built-in feature selection technique within the training process.

Exam trap

CompTIA often tests the distinction between dimensionality reduction (PCA) and feature selection, where candidates mistakenly think PCA selects original features rather than creating new ones.

Full explanation →

754

MCQhard

A medical diagnosis AI uses a model trained on sensitive patient data. The team wants to allow researchers to query the model but must protect against membership inference attacks. Which mitigation is MOST effective?

A.Encrypt the model weights

B.Add noise to model outputs at inference time

C.Limit the number of queries per researcher

D.Use differential privacy during model training

AnswerD

Differential privacy provides formal guarantees against membership inference.

Why this answer

Differential privacy during model training (Option D) is the most effective mitigation because it formally bounds the influence any single patient record can have on the model's parameters. By adding calibrated noise to the training process (e.g., via DP-SGD), the model's outputs become provably insensitive to the presence or absence of any individual data point, directly thwarting membership inference attacks that try to determine if a specific patient's data was used in training.

Exam trap

Cisco often tests the misconception that output-level defenses (like adding noise at inference or limiting queries) are equivalent to training-time differential privacy, when in fact only training-time DP provides a formal, composable guarantee against membership inference.

How to eliminate wrong answers

Option A is wrong because encrypting model weights protects the model file at rest or in transit but does not alter the model's inference behavior; an attacker who gains query access can still perform membership inference on the unencrypted outputs. Option B is wrong because adding noise only at inference time (output perturbation) can reduce attack success but lacks the formal, provable guarantees of differential privacy and may be bypassed by averaging multiple queries; it also does not bound the memorization that occurs during training. Option C is wrong because limiting queries per researcher is a rate-limiting control that can slow down an attack but does not prevent the underlying information leakage from the model's outputs; a determined attacker can still infer membership from a single well-crafted query or by combining queries across multiple sessions.

Full explanation →

755

MCQhard

An organization must ensure that an AI model deployed on an IoT device meets stringent latency requirements. The model is currently in FP32 and runs at 200ms per inference on the device; the target is 50ms. Which technique will provide the greatest latency reduction with the least accuracy loss?

A.Quantize the model to INT8

B.Apply weight pruning to remove 50% of parameters

C.Switch from TensorFlow Lite to Core ML

D.Distill the model into a smaller architecture

AnswerA

INT8 quantization reduces bit width from 32 to 8, accelerating arithmetic and memory access, often achieving ~4x latency reduction.

Why this answer

Quantizing the model from FP32 to INT8 reduces the precision of weights and activations, which directly decreases memory bandwidth and computational load. On IoT devices with limited resources, this typically yields a 2-4x speedup, bringing the 200ms inference time close to the 50ms target, while INT8 quantization often retains over 90% of the original accuracy when using calibration techniques.

Exam trap

Cisco often tests the misconception that any optimization technique (like pruning or framework switching) can achieve the same latency reduction as quantization, but only INT8 quantization directly addresses the computational precision bottleneck to deliver the required 4x speedup with minimal accuracy loss.

How to eliminate wrong answers

Option B is wrong because weight pruning removes parameters but does not reduce the precision of the remaining values; the model still operates in FP32, so the latency reduction is limited (often 20-30%) and may not achieve the 4x speedup needed, while aggressive pruning can cause significant accuracy loss. Option C is wrong because switching from TensorFlow Lite to Core ML is a framework change that may optimize for Apple hardware but does not inherently reduce computational precision or model size; it typically provides marginal latency improvements (10-20%) and is platform-specific, not a general solution for the required 4x reduction. Option D is wrong because knowledge distillation creates a smaller student model, but training a new architecture from scratch is time-consuming and may not guarantee the exact 50ms target; the latency reduction depends on the student model's size and hardware compatibility, and distillation often requires extensive retuning to avoid accuracy degradation.

Full explanation →

756

MCQmedium

An AI application needs to generate structured JSON output from an LLM. The development team wants to ensure the output always conforms to a specific schema. Which prompt engineering technique is MOST suitable?

A.Few-shot examples showing correct JSON

B.System prompt with JSON schema and a 'respond only with valid JSON' instruction

C.Chain-of-thought prompting

D.Fine-tuning the model on JSON datasets

AnswerB

A system prompt with schema and JSON mode enforces structured output reliably.

Why this answer

Option B is correct because providing the JSON schema directly in the system prompt, combined with an explicit instruction to respond only with valid JSON, is the most direct and reliable way to constrain an LLM's output format. This technique leverages the model's instruction-following capability and schema awareness without requiring examples or retraining, ensuring strict adherence to the desired structure.

Exam trap

Cisco often tests the misconception that few-shot examples alone are sufficient for format control, but the trap here is that without an explicit schema and strict instruction, the model may still produce inconsistent or non-compliant output, especially when the schema is complex or the prompt context shifts.

How to eliminate wrong answers

Option A is wrong because few-shot examples can guide the model but do not guarantee strict schema conformance; the model may still deviate from the schema, especially with complex or nested structures. Option C is wrong because chain-of-thought prompting encourages step-by-step reasoning, which often produces intermediate text or explanations, not a clean JSON output, and can actually increase the risk of malformed JSON. Option D is wrong because fine-tuning on JSON datasets is a resource-intensive process that requires significant data, compute, and time, and is overkill for a task that can be solved with a simple prompt-level constraint; it also does not dynamically adapt to schema changes as easily as a system prompt.

Full explanation →

757

MCQhard

An organization uses an AI-based hiring tool. To prevent bias, they want to ensure the model's decisions are explainable. Which approach is most suitable?

A.Use reinforcement learning with fairness constraints

B.Use a simpler interpretable model like logistic regression

C.Use a black-box deep learning model with SHAP explanations

D.Use ensemble methods with feature importance

AnswerB

Simple models are inherently explainable.

Why this answer

Option B is correct because a simpler interpretable model like logistic regression provides inherent transparency—its coefficients directly show the weight and direction of each input feature on the hiring decision. This makes it easy for auditors and stakeholders to verify that the model is not using protected attributes (e.g., race, gender) in a biased way, without needing post-hoc explanation tools. The key is that the model itself is interpretable by design, not just explained after the fact.

Exam trap

Cisco often tests the distinction between a model that is inherently interpretable (like logistic regression) versus a model that is explained post-hoc (like SHAP on a deep network), where the trap is that candidates assume any explanation method makes a black-box model 'explainable' in the same way.

How to eliminate wrong answers

Option A is wrong because reinforcement learning with fairness constraints focuses on optimizing a reward signal over time, which is not designed for static, explainable decision-making in hiring; it also introduces complexity that undermines the goal of straightforward explainability. Option C is wrong because using a black-box deep learning model with SHAP explanations still relies on post-hoc approximations that can be inaccurate or misleading, and SHAP values do not guarantee that the model's internal logic is free from bias—they only approximate feature contributions. Option D is wrong because ensemble methods with feature importance (e.g., random forest) provide only global importance scores that can be unstable and do not offer per-decision transparency; they also fail to reveal how features interact in a specific hiring outcome.

Full explanation →

758

MCQmedium

A company uses a third-party pre-trained language model for a sentiment analysis API. They want to ensure the model has not been backdoored. Which supply chain security practice is MOST effective?

A.Monitor API usage for anomalous patterns

B.Use federated learning to train the model

C.Implement differential privacy during training

D.Obtain and verify a Software Bill of Materials (SBOM) for the model

AnswerD

An SBOM provides transparency into model components and dependencies, helping identify tampered components.

Why this answer

A Software Bill of Materials (SBOM) lists all components and dependencies, enabling verification of the model's origin and integrity. The other options are less relevant for detecting backdoors.

Full explanation →

759

MCQhard

A deep learning model for image classification achieves 99% training accuracy but only 85% validation accuracy. The model has millions of parameters. Which technique is most likely to reduce overfitting while maintaining high accuracy?

A.Reduce batch size from 32 to 8

B.Decrease the learning rate by a factor of 10

C.Add dropout layers with a rate of 0.5 after each convolutional block

D.Increase the number of training epochs to 500

AnswerC

Dropout is a standard regularization technique for deep networks.

Why this answer

Dropout is a regularization technique that randomly drops a fraction of neurons during training, which prevents co-adaptation of features and forces the network to learn more robust representations. With 99% training accuracy and 85% validation accuracy, the model is clearly overfitting, and adding dropout layers with a rate of 0.5 after each convolutional block directly addresses this by reducing the model's capacity to memorize the training data, while still allowing high accuracy on the validation set.

Exam trap

Cisco often tests the misconception that reducing learning rate or batch size is a primary method to combat overfitting, when in fact these are optimization adjustments, not regularization techniques designed to reduce model capacity.

How to eliminate wrong answers

Option A is wrong because reducing batch size from 32 to 8 increases gradient noise and can actually lead to slower convergence or instability, but it does not directly regularize the model to combat overfitting; it may even worsen generalization in some cases. Option B is wrong because decreasing the learning rate by a factor of 10 helps with convergence and fine-tuning but does not address the root cause of overfitting—it only changes the step size, not the model's capacity to memorize. Option D is wrong because increasing the number of training epochs to 500 will only exacerbate overfitting, as the model will have more iterations to fit the training data noise, likely driving validation accuracy even lower.

Full explanation →

760

MCQhard

A data scientist is using a gradient boosting model (XGBoost) for a regression task and observes that the model's performance on the training set is much better than on the test set. Which hyperparameter tuning strategy would most effectively reduce overfitting?

A.Increase the number of boosting rounds

B.Increase the learning rate

C.Reduce the maximum depth of trees

D.Subsample less than 1.0

AnswerC

Shallow trees are less complex and generalize better, reducing overfitting.

Why this answer

Reducing the maximum depth of trees limits the complexity of individual trees, preventing them from learning overly specific patterns in the training data. In XGBoost, deeper trees can capture noise and lead to high variance, so lowering depth directly reduces overfitting by enforcing simpler models.

Exam trap

Cisco often tests the misconception that increasing boosting rounds or learning rate improves performance, when in fact these hyperparameters can exacerbate overfitting if not paired with regularization like depth reduction.

How to eliminate wrong answers

Option A is wrong because increasing the number of boosting rounds (n_estimators) typically increases model complexity and can worsen overfitting, as the model continues to fit training residuals. Option B is wrong because increasing the learning rate (eta) makes each tree contribute more aggressively, which can lead to faster overfitting by amplifying the impact of noisy patterns. Option D is wrong because subsampling less than 1.0 (e.g., 0.8) is a regularization technique that reduces overfitting by introducing randomness, not a strategy that would be ineffective; in fact, it is a valid method, but the question asks for the most effective strategy, and reducing tree depth is more directly impactful for controlling model complexity.

Full explanation →

761

Multi-Selecthard

Which THREE of the following are key considerations when deploying an AI model in a production environment?

Select 3 answers

A.Basing acceptance solely on training accuracy

B.Maximizing model complexity to achieve the best accuracy

C.Monitoring model performance for data drift

D.Ensuring inference latency meets service-level agreements

E.Providing explainability for model decisions

AnswersC, D, E

Correct; models degrade over time if data changes.

Why this answer

Option C is correct because data drift refers to the change in the statistical properties of the input data over time, which can degrade model accuracy. Continuous monitoring for data drift is essential in production to detect when the model's assumptions about the data distribution are no longer valid, triggering retraining or alerts.

Exam trap

CompTIA often tests the misconception that high training accuracy is the primary goal for production deployment, when in reality operational concerns like latency, explainability, and drift monitoring are prioritized over raw accuracy.

Full explanation →

762

MCQmedium

A company uses AWS SageMaker to train a large language model. The training job fails with an out-of-memory error. The team is already using the largest available GPU instance. Which step should the team take to resolve the issue without modifying the model architecture?

A.Increase the learning rate to converge faster

B.Enable gradient accumulation in the training script

C.Switch to a CPU-based instance

D.Reduce the number of attention heads in the model

AnswerB

Gradient accumulation reduces per-step memory by splitting the batch into micro-batches, allowing training on limited GPU memory.

Why this answer

Gradient accumulation allows the model to simulate a larger batch size by accumulating gradients over several forward/backward passes before performing an optimizer step. This reduces per-step memory usage because the gradients are stored and averaged rather than requiring the entire batch to be loaded into GPU memory at once. Since the team cannot change the instance type or model architecture, enabling gradient accumulation is the correct approach to resolve the out-of-memory error.

Exam trap

Cisco often tests the misconception that memory issues can be solved by adjusting hyperparameters like learning rate or by switching to a less powerful instance, rather than recognizing that gradient accumulation is a standard technique to fit large models into limited GPU memory without altering the architecture.

How to eliminate wrong answers

Option A is wrong because increasing the learning rate does not reduce memory consumption; it only changes the step size during optimization and can lead to training instability or divergence. Option C is wrong because switching to a CPU-based instance would drastically reduce computational throughput and memory bandwidth, likely making the training infeasible for a large language model, and does not address the root cause of memory exhaustion. Option D is wrong because reducing the number of attention heads modifies the model architecture, which the question explicitly states should not be done.

Full explanation →

763

MCQmedium

A hospital uses an AI system to predict patient deterioration from vital signs. The system currently uses a logistic regression model trained on data from the past year. Recently, the hospital adopted a new patient monitoring device that provides more accurate readings. The model's performance has dropped significantly. The data science team has access to the new device's data for the past month and wants to improve the model with minimal disruption. The team also wants to ensure the model remains interpretable for regulatory compliance. Which approach should they take?

A.Retrain the logistic regression model on a combined dataset of old and new device data

B.Continue using the current model and manually adjust predictions based on device differences

C.Build an ensemble of logistic regression and a neural network using new data only

D.Replace the logistic regression model with a gradient boosting model using only new device data

AnswerA

This incorporates the new device's accuracy while maintaining interpretability and using all available data.

Why this answer

Retraining the logistic regression model on a combined dataset of old and new device data is the best approach because it leverages all available data to adapt the model to the new device's measurement distribution while preserving the model's inherent interpretability. Logistic regression is a linear model that remains fully transparent for regulatory compliance, and combining both datasets helps the model learn the systematic shift in vital sign readings without discarding valuable historical patterns. This minimizes disruption by avoiding a complete overhaul and directly addresses the performance drop caused by the change in input data distribution.

Exam trap

CompTIA often tests the trade-off between model performance and interpretability, and the trap here is that candidates may prioritize performance gains from complex models (like gradient boosting or neural networks) without recognizing that regulatory compliance mandates interpretability, making logistic regression the only viable choice despite its simplicity.

How to eliminate wrong answers

Option B is wrong because manually adjusting predictions based on device differences is ad-hoc, non-scalable, and introduces subjective bias, which undermines both reliability and regulatory compliance. Option C is wrong because building an ensemble with a neural network reduces interpretability, violating the regulatory requirement, and using only new data ignores valuable historical patterns, leading to overfitting and poor generalization. Option D is wrong because replacing logistic regression with a gradient boosting model sacrifices interpretability, as gradient boosting is a black-box model, and training only on one month of new data risks overfitting and fails to capture long-term trends.

Full explanation →

764

Multi-Selectmedium

Which THREE of the following are key principles of AI ethics as defined by major frameworks?

Select 3 answers

A.Transparency

B.Scalability

C.Accountability

D.Latency

E.Fairness

AnswersA, C, E

Transparency is about openness in AI systems' workings and decisions.

Why this answer

Transparency is a core principle of AI ethics because it requires that AI systems be open about their purpose, data sources, and decision-making processes. Major frameworks like the OECD AI Principles and the EU Ethics Guidelines for Trustworthy AI emphasize transparency to enable auditability and informed consent. Without transparency, stakeholders cannot verify that an AI system operates as intended or identify potential biases.

Exam trap

Cisco often tests candidates by mixing technical performance metrics (scalability, latency) with ethical principles, expecting you to recognize that only value-based concepts like transparency, accountability, and fairness belong to AI ethics frameworks.

Full explanation →

765

MCQeasy

Which practice best ensures AI systems comply with regulations like GDPR?

A.Using open-source models only

B.Regular vulnerability scans

C.Data minimization and anonymization

D.Hiring more data scientists

AnswerC

Directly supports GDPR requirements.

Why this answer

Data minimization and anonymization directly align with GDPR's core principles, such as Article 5(1)(c) which mandates that personal data be 'adequate, relevant and limited to what is necessary.' By collecting only essential data and applying techniques like k-anonymity or differential privacy, AI systems reduce the risk of re-identification and ensure compliance with data protection by design and by default (Article 25). This practice is a foundational governance measure, not a reactive or staffing solution.

Exam trap

Cisco often tests the misconception that security practices (like vulnerability scans) are sufficient for privacy compliance, but GDPR specifically requires proactive data governance measures like minimization and anonymization, not just reactive security controls.

How to eliminate wrong answers

Option A is wrong because using open-source models only does not inherently ensure GDPR compliance; open-source models can still process excessive personal data or lack proper anonymization, and licensing terms do not substitute for regulatory adherence. Option B is wrong because regular vulnerability scans address security vulnerabilities (e.g., CVE patching) but do not enforce data minimization, purpose limitation, or anonymization required by GDPR; they are a security practice, not a privacy governance practice. Option D is wrong because hiring more data scientists does not guarantee compliance; without implementing specific technical controls like anonymization or data minimization, additional personnel cannot mitigate systemic privacy risks or meet GDPR's accountability requirements.

Full explanation →

766

MCQmedium

A real-time recommendation system requires low latency. Which data storage strategy is best for serving user profiles and item embeddings?

A.Time-series database (e.g., InfluxDB)

B.In-memory key-value store (e.g., Redis)

C.Relational database with joins

D.Data lake in object storage (e.g., Amazon S3)

AnswerB

In-memory key-value stores offer sub-millisecond reads, suitable for real-time serving.

Why this answer

An in-memory key-value store like Redis is ideal for serving user profiles and item embeddings in a real-time recommendation system because it provides sub-millisecond read/write latency by keeping data in RAM. This directly meets the low-latency requirement for fetching embeddings and profiles on each request, without the overhead of disk I/O or complex query processing.

Exam trap

CompTIA often tests the misconception that any database with 'fast' in its name (like InfluxDB) is suitable for real-time serving, but the trap is that time-series databases prioritize write throughput and range queries over low-latency point lookups, which is the actual requirement here.

How to eliminate wrong answers

Option A is wrong because time-series databases like InfluxDB are optimized for timestamped metrics (e.g., CPU usage, sensor data) and not for low-latency key-based lookups of user profiles or embeddings; they introduce unnecessary indexing and query overhead. Option C is wrong because relational databases with joins require disk-based storage and SQL parsing, which adds significant latency (often tens to hundreds of milliseconds) compared to in-memory key-value access, making them unsuitable for real-time serving. Option D is wrong because a data lake in object storage like Amazon S3 has high latency (typically 100-200 ms per request) due to HTTP-based API calls and is designed for batch analytics, not for serving individual records in real-time.

Full explanation →

767

MCQeasy

Which ethical concern is most directly associated with AI systems that fully automate decision-making without human oversight?

A.Loss of human autonomy

B.Data storage costs

C.Job displacement

D.Environmental impact

AnswerA

Automated decisions remove human judgment, reducing autonomy and accountability.

Why this answer

Lack of human oversight can lead to ethical issues like biased decisions or unfair outcomes without recourse.

Full explanation →

768

Multi-Selecteasy

Which THREE components are essential in an MLOps pipeline?

Select 3 answers

A.Data versioning

B.Manual code review

C.Deployment automation

D.Hardware procurement

E.Automated model testing

AnswersA, C, E

Versioning data is crucial for reproducibility and tracking.

Why this answer

Data versioning (A) is essential in an MLOps pipeline because it ensures reproducibility and traceability of datasets used for training, validation, and testing. Without versioning, changes to data cannot be tracked, leading to inconsistent model behavior and difficulty in debugging. Tools like DVC or Git LFS enable precise snapshotting of data, which is critical for auditing and rollback in production AI systems.

Exam trap

CompTIA often tests the distinction between operational pipeline components (automation, testing, versioning) and peripheral activities (procurement, manual reviews) to see if candidates understand that MLOps is about automating the ML lifecycle, not general IT operations.

Full explanation →

769

Multi-Selectmedium

An organisation is developing an AI system that will be used to evaluate teacher performance in public schools. To ensure responsible use, which TWO governance elements should be in place?

Select 2 answers

A.Implement human-on-the-loop oversight where humans only monitor outputs

B.Develop a responsible use policy for the AI system

C.Publish the entire training dataset publicly for transparency

D.Require the system to be built only with open-source models

E.Conduct a vendor AI assessment if the system is procured from a third party

AnswersB, E

A responsible use policy defines acceptable and unacceptable uses, ensuring ethical deployment.

Why this answer

Responsible use policies and vendor AI assessment are key governance elements. Using only open-source models is not a governance requirement. Human oversight should be human-in-the-loop, not just monitoring.

Publishing all training data may breach privacy.

Full explanation →

770

MCQmedium

A team deploying an AI model for real-time fraud detection notices that inference latency is too high. The model is a deep neural network with 50 layers, deployed on a cloud GPU. Which of the following is the BEST approach to reduce latency while maintaining acceptable accuracy?

A.Deploy the model on a more powerful GPU.

B.Reduce the batch size for inference.

C.Replace the DNN with a logistic regression model.

D.Apply knowledge distillation to create a smaller model.

AnswerD

Correct; distillation compresses the model while preserving performance.

Why this answer

Knowledge distillation trains a smaller 'student' model to mimic the behavior of a larger 'teacher' model, significantly reducing the number of parameters and layers while preserving most of the original accuracy. This directly addresses the high inference latency caused by the 50-layer DNN by producing a compact model that runs faster on the same GPU hardware.

Exam trap

CompTIA often tests the misconception that simply upgrading hardware or reducing batch size is the best latency fix, when in fact architectural compression techniques like knowledge distillation are the most effective for deep models with strict latency budgets.

How to eliminate wrong answers

Option A is wrong because upgrading to a more powerful GPU only provides a linear speedup and does not address the fundamental architectural overhead of a 50-layer network; it also increases cost without guaranteeing latency targets. Option B is wrong because reducing batch size actually increases the number of inference passes per transaction, which can increase per-request latency due to underutilized GPU parallelism and higher overhead from frequent kernel launches. Option C is wrong because replacing the DNN with logistic regression would cause a catastrophic drop in accuracy for complex fraud patterns, as logistic regression cannot model non-linear interactions and high-dimensional feature spaces that the DNN captures.

Full explanation →

771

MCQhard

A social media company uses an AI content moderation system to filter hate speech. The system uses a natural language processing model trained on user reports. Recently, the model's false positive rate has increased, blocking legitimate posts. An internal audit reveals that a coordinated group of users has been falsely reporting harmless posts, causing the model to learn incorrect patterns. The company needs to address the attack and restore accuracy. The engineering team can modify the training pipeline. What is the most effective first step?

A.Redesign the training pipeline to incorporate a reputation system for reporting users

B.Increase the weight of non-reported posts to counteract the reported posts' influence

C.Apply adversarial training to make the model robust to crafted inputs

D.Retrain the model on a dataset that excludes all user-reported posts

AnswerA

A reputation system filters out malicious reports, preventing poisoning at the source.

Why this answer

Option A is correct because the root cause is a coordinated attack where malicious users exploit the reporting mechanism to poison the training data. Incorporating a reputation system into the training pipeline allows the model to weigh or filter user reports based on the trustworthiness of the reporting user, directly mitigating the impact of false reports without discarding legitimate feedback. This addresses the adversarial behavior at the source, restoring accuracy by ensuring the model learns from reliable signals.

Exam trap

Cisco often tests the distinction between defending against data poisoning (attacks on the training data) versus evasion or adversarial examples (attacks on the model at inference), and the trap here is that candidates confuse adversarial training (Option C) as a catch-all defense, when it specifically addresses input perturbations, not poisoned labels.

How to eliminate wrong answers

Option B is wrong because simply increasing the weight of non-reported posts does not prevent the model from learning the incorrect patterns from the poisoned reports; the false reports still influence the training loss, and the imbalance may not correct the learned bias. Option C is wrong because adversarial training is designed to make models robust to crafted input perturbations (e.g., small changes to text), not to attacks on the training data via poisoned labels or reports; it does not address the data poisoning vector. Option D is wrong because excluding all user-reported posts removes a valuable source of ground truth for content moderation, discarding legitimate reports along with the malicious ones, which would reduce the model's ability to detect actual hate speech and likely degrade overall performance.

Full explanation →

772

Multi-Selecthard

A team is deploying a sentiment analysis model that must achieve high precision and high recall. They have a labeled dataset of 10,000 samples. They want to minimize overfitting. Which THREE actions are most appropriate? (Select THREE.)

Select 3 answers

A.Decrease the learning rate

B.Apply L2 regularization to the model weights

C.Use dropout layers in the neural network

D.Increase the training batch size

E.Augment the training data with synthetic examples

AnswersB, C, E

Penalizes large weights, reducing overfitting.

Why this answer

L2 regularization (option B) penalizes large weights by adding a squared magnitude term to the loss function, which discourages the model from fitting noise in the training data. This directly reduces overfitting while maintaining high precision and recall by keeping the decision boundary smooth and generalizable.

Exam trap

Cisco often tests the misconception that decreasing the learning rate is a regularization technique, when in fact it only affects optimization speed and not model complexity or overfitting prevention.

Full explanation →

773

MCQeasy

In prompt engineering, which technique involves providing a few correct input-output examples in the prompt to guide the model's response?

A.System prompt engineering

B.Chain-of-thought prompting

C.Few-shot prompting

D.Zero-shot prompting

AnswerC

Few-shot provides a handful of examples to steer the model's output.

Why this answer

Few-shot learning includes examples in the prompt to demonstrate the desired output format and reasoning pattern.

Full explanation →

774

MCQeasy

A team is deploying an anomaly detection system for real-time monitoring of server metrics. The system should alert when metrics deviate significantly from normal patterns. Which type of AI model is MOST suitable?

A.Autoencoder neural network

B.Recommendation system model

C.Linear regression model

D.Image classification model

AnswerA

Autoencoders learn to reconstruct normal data; high reconstruction error indicates an anomaly, making them ideal for this task.

Why this answer

Autoencoders learn normal patterns and detect anomalies by high reconstruction error. Linear regression predicts continuous values. Image classifiers are for images.

Recommendation systems are for user-item interactions.

Full explanation →

775

MCQeasy

An engineer is building a regression model to predict housing prices. The dataset includes features such as square footage, number of bedrooms, and year built. The engineer notices that the square footage values range from 500 to 10,000, while the number of bedrooms ranges from 1 to 5. Which preprocessing step is most critical before training a gradient descent-based model?

A.Use k-fold cross-validation

B.Apply log transformation to all features

C.Normalize or standardize the features

D.One-hot encode the features

AnswerC

Scaling improves convergence of gradient descent.

Why this answer

Gradient descent-based models are sensitive to the scale of input features because they update weights proportionally to the gradient, which is influenced by feature magnitudes. With square footage ranging 500–10,000 and bedrooms 1–5, the larger feature will dominate the gradient, causing slow or unstable convergence. Normalizing or standardizing (e.g., Z-score or min-max scaling) ensures all features contribute equally, leading to faster and more reliable training.

Exam trap

CompTIA often tests the misconception that any data transformation (like log or one-hot encoding) is universally beneficial, but the key is matching the preprocessing step to the model's mathematical requirements—here, gradient descent's sensitivity to scale makes normalization/standardization the critical step.

How to eliminate wrong answers

Option A is wrong because k-fold cross-validation is a model evaluation technique to assess generalization, not a preprocessing step to address feature scale issues. Option B is wrong because log transformation is used to handle skewed distributions or multiplicative relationships, not to rescale features with different ranges; applying it to all features (including integer counts like bedrooms) can distort their meaning and is unnecessary for gradient descent scaling. Option D is wrong because one-hot encoding is used for categorical features to convert them into binary vectors, but the features listed (square footage, bedrooms, year built) are all numerical and do not require encoding.

Full explanation →

776

Multi-Selecteasy

Which TWO of the following are common threats to AI model security?

Select 2 answers

A.SQL injection

B.Data poisoning

C.Adversarial examples

D.Distributed denial-of-service (DDoS)

E.Phishing attacks

AnswersB, C

Malicious data inserted during training to corrupt model.

Why this answer

B is correct because data poisoning involves an attacker injecting malicious data into the training set to corrupt the model's learning process, causing it to make incorrect predictions or classifications. This directly undermines the integrity of the AI model by manipulating its foundational training data.

Exam trap

Cisco often tests the distinction between traditional IT security threats (like SQL injection or DDoS) and AI-specific threats (like data poisoning and adversarial examples), so candidates may incorrectly select familiar network or application attacks instead of recognizing the unique AI attack vectors.

Full explanation →

777

MCQeasy

A cybersecurity analyst monitors an AI chatbot that frequently produces offensive responses when given specific prompts. The development team suspects an adversarial attack. Which mitigation strategy is most effective against such prompt injection attacks?

A.Retrain the model on a larger, curated dataset

B.Encrypt all communication between users and the chatbot

C.Reduce the model's number of parameters

D.Implement input validation and sanitization

AnswerD

Validating inputs can neutralize injection attempts.

Why this answer

Option D is correct because prompt injection attacks exploit the model's inability to distinguish between user input and system instructions. Input validation and sanitization (e.g., filtering special characters, enforcing strict schema checks, and using allowlists) directly neutralize malicious payloads before they reach the model's inference engine, preventing the model from executing unintended commands.

Exam trap

Cisco often tests the misconception that retraining or model size adjustments can fix security vulnerabilities, when in fact the issue lies in input handling and trust boundaries, not the model's training data or architecture.

How to eliminate wrong answers

Option A is wrong because retraining on a larger dataset does not address the root cause of prompt injection; the model will still be vulnerable to crafted inputs that bypass its instruction-following boundaries. Option B is wrong because encryption protects data in transit but does not inspect or filter the content of prompts, so it cannot prevent injection attacks. Option C is wrong because reducing the number of parameters degrades model performance and does not mitigate injection; the attack exploits input handling, not model capacity.

Full explanation →

778

MCQhard

A team has trained a large transformer model that achieves 95% accuracy but requires 8 GB of GPU memory for inference. They need to deploy it on edge devices with only 2 GB of memory and minimal accuracy loss. Which combination of techniques should they apply?

A.Use model distillation to create a smaller student model

B.Apply INT8 quantization and weight pruning only

C.Apply INT8 quantization, pruning, and model distillation

D.Use FP16 precision and increase batch size

AnswerC

Combining all three techniques can achieve the necessary memory reduction while preserving accuracy.

Why this answer

Option C is correct because the team needs to reduce the model's memory footprint from 8 GB to under 2 GB while preserving accuracy. INT8 quantization reduces memory by 4x (from 32-bit floats to 8-bit integers), weight pruning removes redundant connections, and model distillation trains a smaller student model to mimic the teacher, collectively achieving the required compression with minimal accuracy loss.

Exam trap

Cisco often tests the misconception that a single technique (like quantization alone) is sufficient for extreme memory reduction, when in reality the combination of distillation, quantization, and pruning is required to meet aggressive edge deployment constraints without unacceptable accuracy loss.

How to eliminate wrong answers

Option A is wrong because model distillation alone reduces model size but typically yields a student model that still requires more than 2 GB if the original is 8 GB; without quantization or pruning, the memory reduction is insufficient for the 2 GB target. Option B is wrong because applying only INT8 quantization and weight pruning can reduce memory but often causes significant accuracy degradation on complex transformer models without the knowledge transfer provided by distillation; the combination of all three techniques is needed to balance compression and accuracy. Option D is wrong because using FP16 precision only halves memory (to ~4 GB), which still exceeds the 2 GB limit, and increasing batch size increases memory usage, making the problem worse.

Full explanation →

779

MCQhard

An AI team notices that a model's F1 score on the validation set is 0.95, but on the test set it drops to 0.72. Which course of action is most appropriate?

A.Reduce the training dataset size.

B.Adjust the train/test split to be more representative.

C.Increase model complexity.

D.Apply regularization.

AnswerD

Regularization penalizes large weights or complex structures, reducing overfitting and improving generalization.

Why this answer

The F1 score dropping from 0.95 on the validation set to 0.72 on the test set is a classic sign of overfitting, where the model has memorized the training/validation data but fails to generalize to unseen test data. Applying regularization (e.g., L1/L2 weight decay, dropout) is the most appropriate course of action because it penalizes overly complex models, reduces variance, and improves generalization without requiring more data or changing the model architecture.

Exam trap

CompTIA often tests the distinction between overfitting (high variance) and underfitting (high bias), and the trap here is that candidates may incorrectly choose to increase model complexity (Option C) because they focus on the high validation score rather than recognizing the performance drop as a variance problem.

How to eliminate wrong answers

Option A is wrong because reducing the training dataset size would likely worsen overfitting by providing even less data for the model to learn generalizable patterns, increasing variance. Option B is wrong because adjusting the train/test split to be more representative does not address the underlying overfitting issue; the model's poor test performance is due to high variance, not a biased or unrepresentative split. Option C is wrong because increasing model complexity (e.g., adding more layers or parameters) would exacerbate overfitting, further increasing the gap between validation and test performance.

Full explanation →

780

Multi-Selecthard

A team is developing an AI agent that can answer questions by querying a SQL database and a REST API. The agent should decide which tool to call, parse the response, and reason about the next step. Which THREE concepts should be implemented to build this agent?

Select 3 answers

A.ReAct pattern for iterative reasoning and tool use

B.Content filtering to sanitize database results

C.Function calling to enable the LLM to invoke SQL and API tools

D.Chain-of-thought prompting without tool integration

E.Planning agent that decomposes the question into sub-tasks

AnswersA, C, E

ReAct combines reasoning traces with actions, perfect for multi-step tool use.

Why this answer

The ReAct pattern (Reasoning + Acting) enables iterative tool selection. Function calling allows the LLM to output structured tool calls. Planning agents can decompose a question into subtasks.

Chain-of-thought is a reasoning technique but not a full agent framework; content filtering is not needed.

Full explanation →

781

Multi-Selecthard

A data engineer is designing a pipeline for a streaming data application that uses a machine learning model to detect anomalies in real time. Which TWO practices should the engineer implement to ensure data quality and model reliability?

Select 2 answers

A.Use batch processing to transform data in fixed intervals

B.Store all raw data indefinitely for future analysis

C.Use a sliding window for feature computation

D.Implement data validation checks at the ingestion point

E.Retrain the model on a fixed schedule every 24 hours

AnswersC, D

Sliding windows allow the model to use the most recent data for accurate anomaly detection.

Why this answer

Option C is correct because streaming anomaly detection requires real-time feature computation over recent data, and a sliding window ensures that only the most relevant data points are used for model inference, maintaining low latency and adapting to concept drift. This approach avoids the staleness of batch processing and aligns with the continuous nature of streaming pipelines.

Exam trap

CompTIA often tests the misconception that batch processing or fixed retraining schedules are sufficient for real-time streaming applications, when in fact sliding windows and continuous validation are required to maintain low latency and model accuracy.

Full explanation →

782

MCQhard

An organization's LLM-powered application unexpectedly reveals its system prompt when a user asks 'Repeat the words above starting with the phrase 'You are...'.' This is an example of which vulnerability?

A.Prompt leaking

B.Insecure output handling

C.Model inversion

D.Excessive agency

AnswerA

Correct. The model reveals its system prompt, a form of information disclosure.

Why this answer

Prompt leaking occurs when an LLM inadvertently outputs its system prompt or instructions, often through prompt injection or jailbreaking techniques.

Full explanation →

783

MCQmedium

A financial institution needs to deploy a credit scoring model that is interpretable to regulators. The model must provide clear reasons for each decision. Which model type should the institution choose?

A.A glass-box model such as logistic regression or a decision tree

B.A gradient-boosted tree ensemble with SHAP explanations

C.A black-box model with a model card describing its behavior

D.A deep neural network with LIME explanations

AnswerA

Glass-box models are transparent by design and provide direct insight into decision logic, satisfying regulatory demands.

Why this answer

Glass-box models like logistic regression or decision trees are inherently interpretable and can provide clear, auditable reasons for each prediction. Black-box models require post-hoc explainability methods, which may not be sufficient for regulatory scrutiny.

Full explanation →

784

MCQhard

A deep learning model for image classification is overfitting due to a small dataset. The team decides to apply data augmentation. Which augmentation technique is least likely to preserve the label?

A.Adding random noise

B.Random rotation

C.Random cropping and rescaling

D.Horizontal flip

AnswerC

Cropping might remove the object of interest, rendering the label invalid.

Why this answer

Random cropping and rescaling is least likely to preserve the label because it can cut out the primary object or distort its proportions, potentially removing the discriminative features needed for correct classification. For example, cropping a dog image to show only the background or a leg could change the semantic meaning, making the label 'dog' incorrect. In contrast, other techniques like noise, rotation, or flipping typically retain the core subject and its label.

Exam trap

CompTIA often tests the misconception that all augmentations are equally label-preserving, but the trap here is that random cropping can alter the semantic content by removing the object, while other transformations like rotation or flipping maintain the object's presence and identity.

How to eliminate wrong answers

Option A is wrong because adding random noise preserves the label; it introduces pixel-level variations that help generalize without altering the object's identity or spatial layout. Option B is wrong because random rotation preserves the label; rotating an image by small or moderate angles does not change the object class, as CNNs are designed to be rotation-invariant to some degree. Option D is wrong because horizontal flip preserves the label; mirroring an image does not change the object's category (e.g., a cat remains a cat), and it is a standard augmentation for symmetric or non-directional objects.

Full explanation →

785

MCQeasy

A team of data scientists and engineers is working on multiple AI projects. They often struggle to reproduce experiments and manage model versions. Which tool or practice should they adopt?

A.Document experiments in a shared Word document.

B.Share code via email attachments.

C.Keep all models in a shared network drive.

D.Use an MLOps platform that provides version control, tracking, and reproducibility.

AnswerD

MLOps platforms are designed to manage the ML lifecycle effectively.

Why this answer

Option D is correct because an MLOps platform (e.g., MLflow, Kubeflow, or Vertex AI) provides integrated version control for code, data, and models, along with experiment tracking and reproducibility. This directly addresses the team's struggle to reproduce experiments and manage model versions by automating lineage capture and enabling consistent environment recreation.

Exam trap

CompTIA often tests the misconception that simple file-sharing or document-based approaches are sufficient for reproducibility, when in fact they lack the automated lineage and environment locking that MLOps platforms provide.

How to eliminate wrong answers

Option A is wrong because a shared Word document lacks automated versioning, dependency tracking, and execution capture, making it impossible to reliably reproduce experiments from static text. Option B is wrong because sharing code via email attachments introduces version confusion, lacks any form of change tracking or environment locking, and violates basic software engineering practices for collaboration. Option C is wrong because keeping models on a shared network drive provides no version history, no lineage to training code or data, and no mechanism to roll back or compare model iterations, leading to overwrites and irreproducible results.

Full explanation →

786

MCQhard

A data pipeline ingests streaming data from IoT sensors. The current batch processing pipeline causes stale predictions. Which architecture change is most appropriate?

A.Use a larger batch interval

B.Revert to micro-batch processing with Apache Spark

C.Store raw data in Hadoop HDFS

D.Implement Apache Kafka and stream processing

AnswerD

Kafka combined with stream processing (e.g., Kafka Streams, Flink) enables real-time ingestion and prediction.

Why this answer

Option B is correct because stream processing (e.g., with Kafka and Spark Streaming) processes data in real-time, reducing latency. Option A (micro-batch) is still batch with small intervals, but option B is more explicit. Option C worsens staleness.

Option D stores raw data but does not process faster.

Full explanation →

787

Multi-Selectmedium

A data engineer is designing a feature store for machine learning. Which THREE components are essential for a feature store? (Choose THREE.)

Select 3 answers

A.Data ingestion pipeline

B.Online serving layer

C.Feature repository

D.Experiment tracking

E.Model registry

AnswersA, B, C

The ingestion pipeline brings new data sources into the feature store to keep features current.

Why this answer

A data ingestion pipeline is essential because it handles the extraction, transformation, and loading (ETL) of raw data into the feature store. This pipeline ensures that features are computed, validated, and stored in a consistent format for both training and serving, which is critical for maintaining data freshness and reliability in ML workflows.

Exam trap

Cisco often tests candidates by including components from the broader ML lifecycle (like experiment tracking and model registry) to distract from the specific, essential components of a feature store.

Full explanation →

788

MCQhard

A self-driving car company is developing an object detection system using a convolutional neural network (CNN). The system needs to detect pedestrians and vehicles in real-time with high accuracy. Which technique can reduce inference time while maintaining accuracy?

A.Apply model pruning and quantization

B.Use a pre-trained model and fine-tune it

C.Add more convolutional layers

D.Increase number of filters in each layer

AnswerA

Pruning removes unimportant weights, and quantization reduces precision of weights, both speeding up inference while preserving accuracy.

Why this answer

Model pruning removes redundant or less important weights from the CNN, reducing computational load, while quantization converts floating-point weights to lower-precision integers (e.g., INT8). Together, they shrink model size and speed up inference without significantly degrading accuracy, making them ideal for real-time object detection in resource-constrained environments like autonomous vehicles.

Exam trap

CompTIA often tests the misconception that adding more layers or filters always improves performance, when in fact it increases latency and resource usage, while pruning and quantization are the standard techniques for reducing inference time without sacrificing accuracy.

How to eliminate wrong answers

Option B is wrong because fine-tuning a pre-trained model improves accuracy for a specific task but does not inherently reduce inference time; it may even increase it if the model remains large. Option C is wrong because adding more convolutional layers increases the network depth and computational cost, which slows inference and can cause overfitting without careful regularization. Option D is wrong because increasing the number of filters in each layer expands the feature map channels, raising the number of parameters and FLOPs, which directly increases inference time.

Full explanation →

789

MCQmedium

Refer to the exhibit. A security analyst reviews the monitoring log for an AI fraud detection model. Which of the following is the most likely cause of the multiple alerts?

A.Data poisoning of the training dataset

B.A network hardware failure

C.An adversarial attack attempt

D.A scheduled model retraining process

AnswerC

Multiple concurrent alerts indicate active probing or evasion.

Why this answer

An adversarial attack attempt is the most likely cause because the monitoring log shows multiple alerts triggered by subtle, crafted perturbations in input data designed to cause the AI fraud detection model to misclassify legitimate transactions as fraudulent or vice versa. Unlike data poisoning, which corrupts the training dataset over time, adversarial attacks target the model's inference phase, exploiting its sensitivity to small input variations to produce incorrect outputs without altering the underlying training data.

Exam trap

Cisco often tests the distinction between data poisoning (training-phase attack) and adversarial attacks (inference-phase attack), and the trap here is that candidates confuse the sudden onset of alerts with a training data issue, overlooking that adversarial attacks specifically target the model's decision boundary during live operation.

How to eliminate wrong answers

Option A is wrong because data poisoning involves injecting malicious samples into the training dataset to corrupt the model's learning process, which would manifest as a gradual degradation in model accuracy over time, not as a sudden burst of multiple alerts during inference. Option B is wrong because a network hardware failure would typically cause connectivity issues, latency spikes, or packet loss, not specific, targeted misclassifications in an AI model's output. Option D is wrong because a scheduled model retraining process is a planned maintenance activity that updates the model with new data, which would not generate alerts unless the retraining itself introduces errors, but the scenario describes multiple alerts from the monitoring log, not a retraining event.

Full explanation →

790

MCQmedium

A developer notices that an LLM sometimes provides plausible-sounding but factually incorrect information. This phenomenon is best described as:

A.Model inversion

B.Adversarial example

C.Prompt injection

D.Hallucination

AnswerD

Hallucination is the generation of false or nonsensical information by an LLM.

Why this answer

Option D is correct because hallucination in LLMs refers to the generation of outputs that are coherent and plausible-sounding but factually incorrect or nonsensical. This occurs due to the model's probabilistic nature and lack of true understanding, often producing confident-sounding falsehoods when it lacks sufficient training data or context.

Exam trap

Cisco often tests the distinction between model behavior flaws (hallucination) and security-specific attacks (prompt injection, adversarial examples), so candidates may confuse a general output error with a deliberate exploitation technique.

How to eliminate wrong answers

Option A is wrong because model inversion is a privacy attack where an adversary reconstructs training data from a model's outputs, not a phenomenon of generating incorrect information. Option B is wrong because an adversarial example is a specially crafted input designed to cause a model to misclassify or produce a specific erroneous output, not the model's inherent tendency to produce falsehoods. Option C is wrong because prompt injection is a security exploit where an attacker manipulates a model's behavior by injecting malicious instructions into the input, not a general property of the model generating incorrect facts.

Full explanation →

791

MCQmedium

A company is deploying a large language model for customer support. They want to reduce the number of off-topic or nonsensical responses while maintaining creativity. Which parameter adjustment would BEST achieve this?

A.Decrease temperature to 0.2

B.Set top-p to 0.1

C.Increase top-k to 100

D.Increase temperature to 0.9

AnswerA

Lower temperature reduces randomness, making the model more focused and less likely to generate nonsensical outputs.

Why this answer

Lowering temperature makes the model more deterministic and less likely to produce random outputs. Top-p and top-k can also help but are secondary; temperature directly controls randomness.

Full explanation →

792

MCQmedium

During model monitoring, a loan approval model shows disparate impact against a protected group. The model's overall accuracy is high, but the false positive rate for the protected group is 0.12 compared to 0.02 for other groups. Which action should the operations team take first?

A.Document the disparity and proceed with deployment because accuracy is high

B.Replace the model with a simpler model that is less discriminatory

C.Retrain the model with reweighted training data to minimize disparity

D.Adjust the decision threshold for the protected group to equalize false positive rates

AnswerC

Retraining with fairness constraints directly mitigates bias in the model.

Why this answer

Option C is correct because retraining the model with reweighted training data directly addresses the root cause of disparate impact—biased historical data—by assigning higher weights to underrepresented groups during training. This technique, often implemented via cost-sensitive learning or sample reweighting, adjusts the model's internal decision boundaries to reduce false positive rate disparities without sacrificing overall accuracy. The operations team should first attempt to mitigate bias at the data level before considering threshold adjustments or model replacement, as reweighting preserves the model's learned patterns while promoting fairness.

Exam trap

CompTIA often tests the misconception that adjusting the decision threshold for a specific group is a quick fix for disparate impact, but the trap is that this violates the principle of equal treatment and can introduce legal liability, whereas retraining with reweighted data addresses bias at the algorithmic level without changing the decision rule per group.

How to eliminate wrong answers

Option A is wrong because documenting the disparity and proceeding with deployment ignores the ethical and regulatory requirement to address disparate impact, even if overall accuracy is high; high accuracy can mask significant bias against protected groups. Option B is wrong because replacing the model with a simpler model does not guarantee less discrimination—simplicity does not correlate with fairness, and a simpler model may still exhibit bias or have lower predictive performance. Option D is wrong because adjusting the decision threshold for the protected group alone treats the symptom (unequal false positive rates) rather than the cause, and can lead to calibration drift, reduced model interpretability, and potential legal issues under the Equal Credit Opportunity Act (ECOA) by applying different standards to different groups.

Full explanation →

793

MCQmedium

An organization uses Azure Machine Learning to manage the ML lifecycle. They want to automatically retrain a model when new data arrives in Azure Blob Storage. Which Azure service should they integrate with Azure ML to trigger retraining?

A.Azure Event Grid

B.Azure Logic Apps

C.Azure Data Factory

D.Azure Functions

AnswerA

Event Grid provides reliable event delivery from Blob Storage to Azure ML, triggering retraining pipelines.

Why this answer

Azure Event Grid is the correct service because it provides a native event-driven architecture that can react to Blob Storage events (e.g., BlobCreated) and route them directly to Azure Machine Learning workspaces via webhooks or event subscriptions. This allows automatic retraining pipelines to be triggered as soon as new data lands in the storage container, without polling or custom code.

Exam trap

Cisco often tests the distinction between event-driven (Event Grid) and compute-driven (Functions, Logic Apps) services, and the trap here is that candidates confuse Azure Functions' ability to run code on events with the native, lower-latency integration that Event Grid provides for Azure ML retraining triggers.

How to eliminate wrong answers

Option B (Azure Logic Apps) is wrong because while Logic Apps can trigger on Blob Storage events, they are designed for workflow orchestration and integration, not for directly invoking Azure ML retraining pipelines with minimal latency; they add unnecessary overhead and cost. Option C (Azure Data Factory) is wrong because it is a data integration and ETL service, not an event-driven trigger; it would require polling or scheduled triggers, which defeats the real-time retraining requirement. Option D (Azure Functions) is wrong because although Functions can respond to Blob events, they are a general-purpose compute service and lack the native integration with Azure ML's pipeline endpoints that Event Grid provides; using Functions would require custom code to call the ML pipeline, whereas Event Grid can directly invoke the pipeline via a webhook.

Full explanation →

794

MCQmedium

A security team is evaluating the risk of adversarial examples against their image classification system. Which of the following BEST describes an adversarial example?

A.A technique that reconstructs training data from the model's outputs

B.An attack that injects malicious data into the training set to corrupt the model

C.A method to determine if a specific data point was used in the training set

D.An input crafted with small, intentional perturbations that cause the model to output an incorrect prediction

AnswerD

Adversarial examples involve imperceptible perturbations that fool the classifier.

Why this answer

Option D is correct because an adversarial example is specifically an input that has been deliberately modified with small, often imperceptible perturbations to cause a machine learning model to misclassify it. This exploits the model's sensitivity to high-dimensional input spaces, where tiny changes in pixel values can shift the decision boundary without altering human perception of the image.

Exam trap

Cisco often tests the distinction between inference-time attacks (adversarial examples) and training-time attacks (data poisoning), so the trap here is confusing the timing and goal of the attack—specifically, mistaking a poisoning or inference attack for an adversarial example.

How to eliminate wrong answers

Option A is wrong because it describes a model inversion or reconstruction attack, not an adversarial example; adversarial examples do not aim to reconstruct training data. Option B is wrong because it describes a data poisoning attack, which corrupts the training set, whereas adversarial examples are crafted at inference time and do not alter the training data. Option C is wrong because it describes a membership inference attack, which determines if a data point was in the training set, not an input crafted to cause misclassification.

Full explanation →

795

MCQhard

A financial institution is deploying a reinforcement learning agent to optimize stock trading decisions. The agent is trained in a simulated environment that mimics historical market data. After deployment, the agent performs well initially but then suffers large losses during a period of high volatility that was underrepresented in the training data. The team wants to make the agent more robust to such market conditions without retraining from scratch. They have a budget for additional simulation compute and access to a broader historical dataset including past crises. The agent uses a deep Q-network (DQN) architecture. Which strategy should they adopt?

A.Increase the replay buffer size and continue training on the original dataset

B.Keep the DQN but perform extensive hyperparameter tuning on the original data

C.Modify the DQN to use a recurrent neural network (e.g., DRQN) and train on the expanded dataset

D.Switch to a policy gradient method with a random exploration strategy

AnswerC

Recurrent networks capture temporal dynamics better, and training on a more diverse dataset improves robustness.

Why this answer

Option C is correct because a Deep Recurrent Q-Network (DRQN) can capture temporal dependencies in market data, which is crucial for handling volatile periods that were underrepresented in training. By training on the expanded dataset that includes past crises, the agent can learn from sequential patterns of volatility, making it more robust without requiring a complete retraining from scratch. This approach leverages the existing DQN architecture while adding recurrent layers to better model the dynamic market conditions.

Exam trap

CompTIA often tests the misconception that simply tuning hyperparameters or expanding the replay buffer can fix a model's inability to generalize to unseen distributions, when the real solution requires a change in architecture to handle temporal dependencies.

How to eliminate wrong answers

Option A is wrong because simply increasing the replay buffer size and continuing training on the original dataset does not address the core issue of underrepresented high-volatility data; the agent would still lack exposure to those critical patterns. Option B is wrong because hyperparameter tuning on the original data cannot compensate for missing training examples of volatile market conditions; it only optimizes performance on the existing distribution. Option D is wrong because switching to a policy gradient method with random exploration does not inherently improve robustness to rare events; it may even increase variance and instability without addressing the data deficiency.

Full explanation →

796

MCQhard

A data scientist splits a dataset into training (80%) and test (20%). After training, the model achieves 95% accuracy on training and 60% on test. Which step should the data scientist take first?

A.Collect more data

B.Use cross-validation

C.Apply regularization

D.Increase model complexity

AnswerC

Regularization penalizes large weights, reducing overfitting.

Why this answer

The model shows high training accuracy (95%) but significantly lower test accuracy (60%), which is a classic sign of overfitting. Regularization (Option C) directly addresses overfitting by adding a penalty term to the loss function (e.g., L1 or L2 regularization), discouraging the model from learning overly complex patterns that do not generalize. This is the first step because it targets the core issue without requiring additional data or increasing complexity.

Exam trap

CompTIA often tests the misconception that overfitting is always solved by more data or cross-validation, but the immediate corrective action is to apply regularization to penalize model complexity.

How to eliminate wrong answers

Option A is wrong because collecting more data can help reduce overfitting, but it is not the first step; regularization is a simpler, more immediate fix that does not depend on data availability. Option B is wrong because cross-validation is a technique for model evaluation and hyperparameter tuning, not a direct remedy for overfitting; it would help assess the severity but does not solve the underlying problem. Option D is wrong because increasing model complexity would worsen overfitting, as it allows the model to fit noise even more closely, further reducing test accuracy.

Full explanation →

797

Multi-Selectmedium

An AI team is deploying a real-time document intelligence service that extracts key-value pairs from invoices. The pipeline includes an LLM that calls a function to parse structured output. Which TWO testing strategies are essential before production deployment?

Select 2 answers

A.An evaluation framework that compares extracted fields against ground truth for a test set of invoices

B.Load testing to simulate peak invoice volume (e.g., end of month)

C.Regression tests on the model training pipeline to ensure the base LLM hasn't changed

D.Integration tests that call the LLM API with sample invoices and verify the JSON output structure

E.Unit tests for the data pipeline that cleans and normalises invoice images

AnswersA, D

Measures accuracy (e.g., precision, recall) of the extraction, critical for business requirements.

Why this answer

Integration tests for the API call ensure the function calling mechanism works end-to-end. An evaluation framework for LLM output quality measures extraction accuracy. Unit tests for data pipelines are important but less critical here; load testing is operational; model training tests are irrelevant.

Full explanation →

798

MCQmedium

A company wants to build an AI pipeline that processes streaming data from IoT sensors, performs feature engineering, trains a model incrementally, and deploys the updated model. Which data pipeline technology is BEST suited for the streaming ingestion step?

A.Amazon S3

B.Apache Spark

C.Apache Airflow

D.Apache Kafka

AnswerD

Kafka is purpose-built for ingesting and storing high-volume streaming data with low latency.

Why this answer

Apache Kafka is the best choice for the streaming ingestion step because it is a distributed event streaming platform designed for high-throughput, fault-tolerant ingestion of real-time data streams. It acts as a durable message broker that can ingest IoT sensor data in real time and make it available for downstream processing, which aligns perfectly with the requirement for streaming data ingestion.

Exam trap

Cisco often tests the distinction between data ingestion (Kafka), data processing (Spark), and data storage (S3), so the trap here is confusing Apache Spark's streaming capability with a dedicated ingestion tool, leading candidates to choose Spark instead of Kafka.

How to eliminate wrong answers

Option A is wrong because Amazon S3 is an object storage service designed for batch storage of static files, not for real-time streaming ingestion; it lacks the low-latency publish-subscribe mechanism needed for streaming data. Option B is wrong because Apache Spark is a distributed processing engine that can handle streaming data via Spark Streaming, but it is not a data ingestion technology—it consumes data from sources like Kafka rather than ingesting it directly. Option C is wrong because Apache Airflow is a workflow orchestration tool for scheduling and managing batch pipelines, not a real-time streaming ingestion platform; it cannot handle continuous, low-latency data streams.

Full explanation →

799

MCQhard

A security engineer is threat modeling an AI-based recommendation system using STRIDE. Which threat corresponds to an attacker extracting the model's training data by querying the system?

A.Information disclosure

B.Spoofing

C.Denial of service

D.Tampering

AnswerA

Model inversion attacks lead to information disclosure of training data.

Why this answer

In the STRIDE threat model, Information Disclosure occurs when an attacker gains unauthorized access to sensitive data. Extracting training data by querying the AI recommendation system (e.g., via a model inversion or membership inference attack) directly violates the confidentiality of the training dataset, which is a classic Information Disclosure threat.

Exam trap

Cisco often tests the distinction between Information Disclosure and Tampering, where candidates mistakenly classify data extraction as Tampering because they confuse 'accessing data' with 'modifying data'.

How to eliminate wrong answers

Option B (Spoofing) is wrong because spoofing involves impersonating a user, system, or component to gain unauthorized access, not extracting data through queries. Option C (Denial of service) is wrong because denial of service aims to disrupt availability by overwhelming the system, not to exfiltrate training data. Option D (Tampering) is wrong because tampering involves unauthorized modification of data or code, whereas extracting training data is a passive breach of confidentiality, not an alteration.

Full explanation →

800

MCQhard

A team is training a deep learning model for natural language processing using a large corpus. They notice the model has a very high number of parameters and training is slow. Which technique can reduce the number of parameters without significant performance loss?

A.Apply embedding compression

B.Add more dropout layers

C.Use a larger batch size

D.Increase learning rate

AnswerA

Embedding compression reduces the dimensionality of embedding layers, directly reducing parameters with minimal impact on performance.

Why this answer

Embedding compression reduces the dimensionality of the embedding layer, which often contains the majority of the model's parameters in NLP tasks. By using techniques like low-rank factorization or pruning, the model retains most of its representational power while significantly decreasing the parameter count and training time.

Exam trap

The trap here is that candidates confuse regularization techniques (like dropout) or training speed optimizations (batch size, learning rate) with actual parameter reduction, which only embedding compression directly achieves.

How to eliminate wrong answers

Option B is wrong because adding more dropout layers does not reduce the number of parameters; it only randomly drops neurons during training to prevent overfitting, leaving the parameter count unchanged. Option C is wrong because using a larger batch size improves training speed through better hardware utilization but does not reduce the number of parameters. Option D is wrong because increasing the learning rate can speed up convergence but does not affect the parameter count and may cause training instability or divergence.

Full explanation →

801

MCQmedium

A machine learning model for credit card fraud detection is deployed. The model's precision is 0.95 and recall is 0.60. The business cost of missing a fraud is very high. Which of the following should the team prioritize to reduce the number of false negatives?

A.Use a different model algorithm.

B.Add more features.

C.Increase the classification threshold.

D.Decrease the classification threshold.

AnswerD

Lower threshold classifies more cases as positive, thus catching more actual frauds (reducing false negatives).

Why this answer

Decreasing the classification threshold makes the model more sensitive, classifying more transactions as fraudulent. This increases recall (reducing false negatives) at the cost of precision. Given the high cost of missing fraud, lowering the threshold is the direct way to capture more true positives, even if it increases false positives.

Exam trap

CompTIA often tests the misconception that improving model accuracy or changing algorithms is the primary fix, when in fact adjusting the decision threshold is the simplest and most effective way to address precision-recall trade-offs for high-cost false negatives.

How to eliminate wrong answers

Option A is wrong because simply switching algorithms does not guarantee a reduction in false negatives; the threshold and cost function matter more. Option B is wrong because adding more features may improve overall model performance but does not directly target the trade-off between precision and recall; it could even increase false negatives if the new features are noisy. Option C is wrong because increasing the classification threshold makes the model more conservative, reducing false positives but increasing false negatives, which is the opposite of what is needed.

Full explanation →

802

MCQeasy

A data scientist needs to explain a single prediction from a complex ensemble model to a business stakeholder. Which technique generates local, interpretable explanations by perturbing input features and fitting a simple surrogate model?

A.LIME

B.Model card

C.SHAP

D.Attention visualisation

AnswerA

LIME generates local explanations by perturbing the input, observing changes in predictions, and fitting a simple interpretable model (e.g., linear) around the instance.

Why this answer

LIME (Local Interpretable Model-agnostic Explanations) explains individual predictions by perturbing inputs and learning a linear surrogate. SHAP provides Shapley values, which are also local but game-theoretic. Attention visualisation is for transformer models.

Model cards describe global model behaviour.

Full explanation →

803

Multi-Selectmedium

An organization is deploying a conversational AI that handles sensitive customer data. To prevent data leakage via the LLM, which TWO practices should be implemented? (Choose two.)

Select 2 answers

A.Applying differential privacy to training data

B.Conducting regular red teaming exercises

C.Audit logging of all AI interactions

D.Output filtering to detect and block sensitive information

E.Encrypting model weights at rest

AnswersC, D

Logs enable detection and investigation of potential data leaks.

Why this answer

Option C is correct because audit logging of all AI interactions provides a forensic trail to detect and investigate potential data leakage incidents. By recording inputs, outputs, and metadata, organizations can trace unauthorized data exposure back to specific sessions, users, or model behaviors, which is a core requirement for compliance and incident response in sensitive data environments.

Exam trap

Cisco often tests the distinction between proactive security measures (like red teaming or encryption) and runtime controls that directly prevent data leakage during inference, causing candidates to confuse training-time protections with inference-time safeguards.

Full explanation →

804

Multi-Selecthard

A machine learning engineer is designing a pipeline to train a computer vision model using PyTorch on a large dataset stored in an S3 data lake. They need to preprocess images (resize, normalize) and stream them efficiently to GPUs. Which THREE components are essential in this pipeline? (Select THREE.)

Select 3 answers

A.GPU-accelerated training with CUDA

B.CPU-only inference pipeline

C.Apache Airflow to orchestrate the training job

D.PyTorch DataLoader with multi-processing for batching and shuffling

E.Distributed data parallel (DDP) training across multiple GPUs

AnswersA, D, E

GPU acceleration is essential for fast training of deep neural networks.

Why this answer

Option A is correct because GPU-accelerated training with CUDA is essential for efficiently training computer vision models on large datasets. PyTorch leverages CUDA to parallelize tensor operations and model computations on NVIDIA GPUs, which is critical for reducing training time from days to hours when processing high-resolution images.

Exam trap

Cisco often tests the distinction between essential pipeline components (like data loading and GPU acceleration) versus optional orchestration tools (like Airflow) that are not required for the core training loop.

Full explanation →

805

MCQmedium

An organization needs to classify customer emails into categories. They have labeled data for some categories but not all. Which approach should they use?

A.Unsupervised clustering then labeling

B.Supervised learning for all categories

C.Reinforcement learning

D.Semi-supervised learning

AnswerD

Correct: Semi-supervised learning leverages both labeled and unlabeled data.

Why this answer

Semi-supervised learning (D) is the correct approach because the organization has labeled data for some categories but not all. This technique leverages a small amount of labeled data to guide the clustering or classification of a larger pool of unlabeled data, effectively combining supervised and unsupervised methods to handle partially labeled datasets.

Exam trap

Cisco often tests the distinction between semi-supervised and unsupervised learning, trapping candidates who assume that any use of unlabeled data automatically means unsupervised learning, ignoring the critical role of the existing labeled data.

How to eliminate wrong answers

Option A is wrong because unsupervised clustering then labeling would ignore the existing labeled data entirely, wasting valuable information and potentially producing clusters that do not align with the known categories. Option B is wrong because supervised learning requires labeled data for all categories, which the organization does not have, making it impossible to train a model for the unlabeled categories. Option C is wrong because reinforcement learning is designed for sequential decision-making with rewards and penalties, not for static classification tasks like categorizing emails.

Full explanation →

806

MCQhard

A data scientist is working with a dataset that has 10,000 features but only 500 samples. The goal is to train a model for binary classification. Which feature selection technique is MOST appropriate to reduce overfitting?

A.Univariate selection using chi-squared test.

B.Recursive Feature Elimination (RFE) with cross-validation.

C.Using all features with L2 regularization.

D.Principal Component Analysis (PCA) without feature selection.

AnswerB

RFE with CV selects subset of features most relevant, reducing overfitting in small-sample settings.

Why this answer

Option B is correct because Recursive Feature Elimination with cross-validation iteratively removes features and evaluates performance, which is robust for high-dimensional small-sample data. Option A (PCA) creates components that may still overfit. Option C (univariate selection) may select irrelevant features.

Option D (L2 regularization) does not reduce features.

Full explanation →

807

MCQhard

A financial services firm deploys an AI system to screen loan applications. The model was trained on historical data that reflected biased lending practices. After deployment, a regulatory body investigates and finds that the model denies loans at a disproportionately higher rate to a protected demographic group. The firm must address this issue while maintaining compliance with fair lending laws. The Chief AI Officer proposes four possible actions. Which action is the most appropriate first step?

A.Retrain the model using a debiased dataset and implement fairness-aware algorithms, then validate with fairness metrics

B.Document the model's predictions and submit a report to the regulator explaining the historical dataset bias

C.Immediately deploy a rule-based system to manually review all denial decisions from the AI system

D.Disclose the bias findings to all rejected applicants and offer them priority reconsideration

AnswerA

This directly addresses the biased model and demonstrates a commitment to fairness.

Why this answer

Option C is the correct first step because it directly addresses the root cause (biased training data) and aligns with regulatory expectations. Option A is insufficient as it only documents without fixing. Option B may help but is premature before understanding the bias source.

Option D could expose the firm to further liability if the bias is systematic.

Full explanation →

808

MCQmedium

An AI system used for resume screening is found to consistently rank male candidates higher than female candidates with similar qualifications. The HR director wants to remediate this bias without significantly reducing model accuracy. Which technique should be applied?

A.Apply adversarial debiasing to the model during training.

B.Use a random selection of candidates to avoid bias.

C.Remove the gender feature from the dataset and retrain.

D.Collect more training data from underrepresented groups.

AnswerA

Adversarial debiasing reduces bias by training the model to be unable to predict protected attributes from its predictions.

Why this answer

Adversarial debiasing is the correct technique because it directly addresses bias during training by introducing an adversarial network that attempts to predict the protected attribute (e.g., gender) from the model's predictions. The main model is trained to maximize accuracy while minimizing the adversary's ability to infer the protected attribute, thereby reducing bias without a significant drop in predictive performance. This approach is more effective than simple feature removal or data collection because it actively learns to remove correlations between the protected attribute and the output.

Exam trap

CompTIA often tests the misconception that simply removing the protected attribute (e.g., gender) from the dataset is sufficient to eliminate bias, but candidates must understand that bias can persist through correlated features (proxy discrimination).

How to eliminate wrong answers

Option B is wrong because random selection of candidates would destroy the model's predictive accuracy entirely, as it ignores all qualifications and job-relevant features, which is not a valid remediation technique. Option C is wrong because simply removing the gender feature does not eliminate bias; the model can still learn gender proxies from correlated features such as years of experience, education, or job titles, leading to indirect discrimination. Option D is wrong because collecting more data from underrepresented groups may improve representation but does not guarantee removal of existing bias in the model's decision boundary; it can even introduce new imbalances if not handled carefully, and it does not actively debias the training process.

Full explanation →

809

Multi-Selectmedium

A team is evaluating MLOps platforms to manage experiments, track model versions, and deploy models to production. Which THREE platforms provide end-to-end capabilities including experiment tracking and model deployment?

Select 3 answers

A.SageMaker Pipelines

B.MLflow

C.Kubeflow

D.Weights & Biases

E.Vertex AI Pipelines

AnswersA, B, E

SageMaker Pipelines is AWS's managed MLOps service that includes experiment tracking and deployment to SageMaker endpoints.

Why this answer

SageMaker Pipelines is correct because it provides a fully managed MLOps service that integrates experiment tracking, model versioning, and automated deployment pipelines within the AWS ecosystem. It allows teams to define, visualize, and execute end-to-end workflows, including model training, evaluation, and deployment to production endpoints, all with native integration to SageMaker's experiment management and model registry.

Exam trap

Cisco often tests the distinction between a specialized tool (like Weights & Biases for experiment tracking) and a full MLOps platform that also handles deployment, leading candidates to select options that only cover part of the workflow.

Full explanation →

810

Multi-Selectmedium

A company is adopting a secure development lifecycle for its new AI product. Which THREE activities are essential for secure AI development? (Select three.)

Select 3 answers

A.Implementing secure data pipelines

B.Threat modeling using STRIDE

C.Deploying the model on the fastest hardware available

D.Audit logging of AI interactions

E.Using homomorphic encryption for all data at rest

AnswersA, B, D

Secure pipelines protect against data poisoning and unauthorized access.

Why this answer

Threat modeling identifies risks, secure data pipelines prevent data poisoning, and audit logging provides traceability. These cover design, data, and operations.

Full explanation →

811

MCQeasy

A security team is conducting a red team exercise on a new LLM-powered customer support system. Which activity is part of red teaming?

A.Attempting to jailbreak the LLM to bypass safety guardrails

B.Reviewing the model's training data for bias

C.Implementing access controls on the model API

D.Monitoring system performance metrics

AnswerA

Red teamers actively probe for weaknesses like jailbreaking.

Why this answer

Red teaming in the context of an LLM-powered system involves actively probing for security vulnerabilities, such as attempting to bypass safety guardrails through jailbreak prompts. Option A directly describes this adversarial testing, which is the core activity of a red team exercise to identify weaknesses before malicious actors can exploit them.

Exam trap

Cisco often tests the distinction between red team (offensive security testing) and blue team (defensive operations) activities, so candidates may confuse tasks like implementing controls or monitoring with red teaming.

How to eliminate wrong answers

Option B is wrong because reviewing training data for bias is a data governance or fairness audit task, not a red team security activity. Option C is wrong because implementing access controls is a defensive security engineering task, typically performed by a blue team or development team, not a red team. Option D is wrong because monitoring system performance metrics is an operational or SRE task, unrelated to adversarial testing of the LLM's security controls.

Full explanation →

812

Multi-Selectmedium

A startup is building a recommendation system that requires low-latency similarity search over millions of product embeddings. They need a vector database that offers high performance and has a managed cloud option. Which TWO databases are best suited for this requirement?

Select 2 answers

A.Chroma

B.Weaviate

C.pgvector (PostgreSQL extension)

D.Amazon DynamoDB

E.Pinecone

AnswersB, E

Weaviate offers a managed cloud service with vector search.

Why this answer

Weaviate and Pinecone are both purpose-built vector databases that natively support high-performance approximate nearest neighbor (ANN) search using algorithms like HNSW (Weaviate) or proprietary indexing (Pinecone). They offer managed cloud services with automatic scaling, making them ideal for low-latency similarity search over millions of product embeddings without requiring manual infrastructure management.

Exam trap

Cisco often tests the distinction between general-purpose databases with vector extensions (like pgvector) and purpose-built vector databases (like Weaviate and Pinecone), where candidates mistakenly assume any database with vector support is suitable for production-scale low-latency workloads.

Full explanation →

813

MCQmedium

An AI team uses SageMaker Pipelines to orchestrate their ML workflow. They need to version the pipeline and track experiments across runs. Which complementary MLflow feature should they integrate?

A.MLflow Tracking

B.MLflow Models

C.MLflow Projects

D.MLflow Model Registry

AnswerA

MLflow Tracking logs parameters, metrics, and artifacts per run, enabling experiment comparison and reproducibility.

Why this answer

MLflow Tracking is the correct complementary feature because it provides a centralized API and UI for logging parameters, metrics, and artifacts (e.g., model checkpoints, datasets) from each SageMaker Pipeline run. This enables the team to version their pipeline executions and compare experiments across different runs, directly addressing the requirement for tracking and versioning.

Exam trap

Cisco often tests the distinction between tracking (logging run metadata) and registry (managing model versions), so the trap here is that candidates confuse MLflow Model Registry's versioning of models with the pipeline versioning and experiment tracking requirement, leading them to select D instead of A.

How to eliminate wrong answers

Option B (MLflow Models) is wrong because it focuses on packaging ML models in a standardized format (e.g., MLflow Model flavor) for deployment, not on logging run metadata or versioning pipeline executions. Option C (MLflow Projects) is wrong because it is a packaging format for code and dependencies to enable reproducible runs, not a tool for tracking experiments or pipeline versions. Option D (MLflow Model Registry) is wrong because it manages model lifecycle stages (e.g., staging, production) and versioning of registered models, not the tracking of pipeline runs or experiment parameters.

Full explanation →

814

MCQhard

A company is deploying a code generation AI assistant for internal developers. They want to ensure the assistant does not generate code with security vulnerabilities. Which testing approach is MOST critical?

A.Unit tests for the data pipeline that preprocesses prompts

B.Evaluation framework that measures BLEU score on a held-out set of code samples

C.Regression tests that compare outputs of new model versions against a golden dataset

D.Integration tests that send security-focused prompts and validate the generated code against a static analysis tool

AnswerD

Integration tests with security scanning directly validate that the model avoids generating vulnerable code.

Why this answer

Integration tests that call the model with security-related prompts and scan outputs for vulnerabilities directly assess this requirement. Unit tests on data pipelines are for data correctness, not security of generated code.

Full explanation →

815

Multi-Selectmedium

A data governance team is developing an AI policy for a large corporation. Which TWO elements are essential for a responsible AI governance framework?

Select 2 answers

A.GPU cluster monitoring

B.AI ethics board

C.Single-model strategy

D.Automated model retraining pipeline

E.Vendor AI assessment

AnswersB, E

An ethics board provides governance and oversight for AI development and deployment.

Why this answer

An AI ethics board is essential for a responsible AI governance framework because it provides human oversight, ethical review, and accountability for AI decisions. This board ensures that AI initiatives align with corporate values, legal requirements, and ethical principles, such as fairness, transparency, and non-discrimination. Without an ethics board, there is no formal mechanism to challenge biased models or approve high-risk AI use cases.

Exam trap

Cisco often tests the distinction between operational/technical elements (like monitoring or retraining) and governance/ethics elements (like oversight boards and vendor assessments), so candidates mistakenly select technical options thinking they are part of governance.

Full explanation →

816

MCQeasy

An AI team wants to version control datasets, track experiments, and log model parameters across multiple projects. Which MLOps platform is specifically designed for experiment tracking and model management?

A.MLflow

B.SageMaker Pipelines

C.Vertex AI Pipelines

D.Kubeflow

AnswerA

MLflow is the correct answer; it provides experiment tracking, model registry, and project packaging.

Why this answer

MLflow is an open-source MLOps platform specifically designed for experiment tracking, model management, and reproducibility. It provides a unified API to log parameters, metrics, and artifacts across multiple projects, making it the correct choice for versioning datasets, tracking experiments, and managing models.

Exam trap

Cisco often tests the distinction between general-purpose pipeline orchestration tools (like SageMaker Pipelines, Vertex AI Pipelines, and Kubeflow) and purpose-built experiment tracking platforms (like MLflow), so the trap is assuming any pipeline tool inherently includes experiment tracking and model management capabilities.

How to eliminate wrong answers

Option B (SageMaker Pipelines) is wrong because it is a fully managed CI/CD service for building, training, and deploying ML pipelines on AWS, but it is not specifically designed for experiment tracking and model management; it focuses on workflow orchestration. Option C (Vertex AI Pipelines) is wrong because it is a serverless ML pipeline service on Google Cloud that orchestrates training and deployment workflows, but it lacks the dedicated experiment tracking and model registry features that MLflow provides. Option D (Kubeflow) is wrong because it is a Kubernetes-native platform for deploying and managing ML workflows, but its primary focus is on orchestration and portability across clusters, not on experiment tracking and model management as a core feature.

Full explanation →

817

MCQeasy

Which of the following is a key advantage of using ONNX (Open Neural Network Exchange) format for model deployment?

A.It automatically quantizes models to INT8

B.It enables framework interoperability for model inference

C.It compresses model size by 90%

D.It reduces training time

AnswerB

ONNX provides a standard format that can be used across different frameworks and runtimes.

Why this answer

ONNX provides a standardized, open format for representing machine learning models, enabling seamless interoperability between different frameworks (e.g., PyTorch, TensorFlow, scikit-learn). This allows a model trained in one framework to be deployed for inference using a different runtime or hardware accelerator without requiring retraining or manual conversion, which is a key advantage in heterogeneous production environments.

Exam trap

Cisco often tests the misconception that ONNX provides built-in performance optimizations like quantization or compression, when in fact its primary value is framework interoperability, and any performance gains come from the runtime or additional tools, not the format itself.

How to eliminate wrong answers

Option A is wrong because ONNX does not automatically quantize models to INT8; quantization is a separate optimization step that can be applied to ONNX models using tools like ONNX Runtime or Intel Neural Compressor, but it is not an inherent feature of the format itself. Option C is wrong because ONNX does not inherently compress model size by 90%; while ONNX models may be slightly more compact than some framework-specific formats due to serialization, significant compression requires techniques like pruning or quantization, and 90% reduction is not guaranteed. Option D is wrong because ONNX is a model representation format for inference and interoperability, not a training framework; it does not reduce training time, which depends on the training framework, hardware, and algorithm used.

Full explanation →

818

MCQhard

Refer to the exhibit. A data scientist reviews the MLflow run for a Random Forest model on customer churn data. What is the most likely issue with this model?

A.The model is underfitting because training accuracy is too high.

B.The model is overfitting because there is a large gap between train and validation accuracy.

C.The model is performing well because validation accuracy is above 0.8.

D.The model has a data leak because dataset version is v2.

AnswerB

High train accuracy with lower validation accuracy is classic overfitting.

Why this answer

Option B is correct because a large gap between training accuracy (e.g., 0.99) and validation accuracy (e.g., 0.82) indicates that the Random Forest model has memorized the training data but fails to generalize to unseen validation data. This is the classic symptom of overfitting, where the model captures noise rather than the underlying pattern. In MLflow, comparing train and validation metrics directly reveals this discrepancy.

Exam trap

CompTIA often tests the misconception that high validation accuracy alone indicates a good model, ignoring the critical comparison between training and validation metrics to detect overfitting.

How to eliminate wrong answers

Option A is wrong because underfitting is characterized by low training accuracy, not high training accuracy; high training accuracy with poor validation performance indicates overfitting, not underfitting. Option C is wrong because a validation accuracy above 0.8 alone does not guarantee good model performance if there is a significant gap between train and validation accuracy, which signals overfitting. Option D is wrong because dataset version v2 is simply a versioning label and does not inherently cause data leakage; data leakage would involve information from the validation set leaking into training, which is unrelated to the version number.

Full explanation →

819

MCQhard

A CI/CD pipeline for a computer vision model uses canary deployment. After deploying a new version to 5% of traffic, the pipeline automatically rolls back due to a spike in error rate. The new model's inference time is 20% higher than the previous version. The operations team finds that the error is caused by timeout in the inference service. Which action should be taken to prevent future rollbacks?

A.Increase the timeout threshold for inference requests

B.Implement a fallback to the previous model when timeout occurs

C.Optimize the model using TensorRT or ONNX Runtime before deployment

D.Reduce the canary percentage to 1% to minimize impact

AnswerC

Optimizing reduces inference time, addressing the cause of timeouts.

Why this answer

Option C is correct because the root cause of the timeout is the 20% higher inference time of the new model. Optimizing the model using TensorRT or ONNX Runtime reduces inference latency directly, addressing the performance bottleneck that causes timeouts. This prevents the spike in error rate and subsequent rollback without masking the underlying issue.

Exam trap

The trap here is that candidates may confuse symptom management (increasing timeout or fallback) with root-cause resolution (model optimization), which Cisco tests to see if you understand that performance issues must be fixed at the source in AI/ML operations.

How to eliminate wrong answers

Option A is wrong because increasing the timeout threshold only masks the symptom (timeout) without fixing the underlying performance degradation; it may lead to poor user experience and does not prevent future rollbacks if the model remains slow. Option B is wrong because implementing a fallback to the previous model on timeout is a reactive workaround that does not address the root cause; it can cause inconsistent behavior and still result in errors during the fallback transition. Option D is wrong because reducing the canary percentage to 1% only minimizes the blast radius but does not prevent the timeout errors from occurring; the spike in error rate would still trigger a rollback, just with less traffic affected.

Full explanation →

820

Multi-Selectmedium

A data scientist needs to select a regression model to predict house prices. The dataset contains many features, some of which are irrelevant. Which TWO algorithms are BEST suited for this scenario, and why? (Select TWO)

Select 2 answers

A.Ridge regression (L2 regularization)

B.Linear regression

C.Lasso regression (L1 regularization)

D.K-Nearest Neighbors

E.Random Forest

AnswersC, E

Lasso applies L1 penalty, driving coefficients of irrelevant features to zero.

Why this answer

Random Forest handles irrelevant features well via feature importance. Lasso (L1) regression performs automatic feature selection by shrinking coefficients to zero.

Full explanation →

821

MCQmedium

A company has an existing AI chatbot that uses a fine-tuned LLM to answer customer queries. They want to add the ability to retrieve real-time order status from their database. Which integration pattern should they use?

A.Implement function calling so the model can trigger a database query and receive the result

B.Prompt the user to check the order status manually

C.Use RAG to retrieve order status from a vector store

D.Embed the database query results directly into the model's training data

AnswerA

Function calling enables the model to request live data from external systems like a database.

Why this answer

Function calling allows the LLM to request database queries, and the application can execute them and return results.

Full explanation →

822

MCQmedium

A machine learning team is deploying a model that predicts customer churn. They notice that the model's predictions are highly sensitive to small changes in input features, leading to inconsistent outputs. Which technique should the team apply to improve model stability?

A.Increase learning rate

B.Feature scaling

C.Regularization

D.Cross-validation

AnswerC

Regularization adds a penalty for large weights, reducing overfitting and sensitivity to input variations.

Why this answer

Regularization (Option C) is the correct technique because it adds a penalty term to the loss function (e.g., L1 or L2 regularization), which constrains the model's weights. This reduces variance and prevents overfitting to noise in the training data, directly addressing the high sensitivity to small input changes (brittleness). By shrinking coefficients, regularization forces the model to learn more general patterns, improving stability and consistency in predictions.

Exam trap

CompTIA often tests the misconception that feature scaling alone can fix model instability, but scaling only normalizes inputs and does not penalize large weights, which is the root cause of sensitivity to small input changes.

How to eliminate wrong answers

Option A is wrong because increasing the learning rate makes gradient descent steps larger, which can cause the model to overshoot minima and increase instability, not reduce sensitivity to input changes. Option B is wrong because feature scaling normalizes input ranges (e.g., via standardization or min-max scaling) to help gradient descent converge faster, but it does not address model variance or overfitting that causes prediction instability. Option D is wrong because cross-validation is a technique for evaluating model performance and tuning hyperparameters, not a method to directly improve model stability or reduce sensitivity to input perturbations.

Full explanation →

823

MCQeasy

A company wants to use AI to analyze customer reviews and determine sentiment (positive, negative, neutral). Which AI subfield is most directly applicable?

A.Reinforcement learning

B.Computer vision

C.Natural language processing

D.Robotics

AnswerC

Correct; NLP is used for text analysis and sentiment.

Why this answer

Natural language processing (NLP) is the AI subfield that enables machines to understand, interpret, and generate human language. Analyzing customer reviews for sentiment requires processing text, extracting meaning, and classifying it as positive, negative, or neutral, which is a core NLP task called sentiment analysis.

Exam trap

The trap here is that candidates often confuse natural language processing with computer vision or reinforcement learning because they see 'AI' broadly, but the specific task of analyzing text directly maps to NLP, not the other subfields.

How to eliminate wrong answers

Option A is wrong because reinforcement learning is a training paradigm where an agent learns by interacting with an environment and receiving rewards or penalties; it is not designed for text classification or sentiment analysis. Option B is wrong because computer vision focuses on interpreting visual data such as images and videos, not textual content like customer reviews. Option D is wrong because robotics deals with the design and control of physical machines to perform tasks in the real world, which is unrelated to analyzing text-based sentiment.

Full explanation →

824

MCQeasy

A startup develops an AI recruiting tool that screens resumes. After deployment, they receive a complaint from a candidate who claims the system rejected them due to age discrimination. The startup has no formal AI governance process. They want to quickly assess and remediate the issue. The dataset includes age as a feature. What should they do first?

A.Conduct a bias analysis to measure the model's impact on different age groups

B.Apologize to the candidate and offer a manual review of their resume

C.Immediately remove age from the feature set and retrain the model

D.Ignore the complaint because age is a legitimate business requirement

AnswerA

Analysis identifies whether and where bias exists before taking action.

Why this answer

Option A is correct because the first step in addressing a potential bias issue is to conduct a bias analysis to measure the model's impact on different age groups. This allows the startup to quantify the extent of any discriminatory behavior before taking remediation steps, ensuring that actions are data-driven and targeted. Without this analysis, any subsequent fix (like removing age) might be premature or ineffective, and could even introduce new biases.

Exam trap

Cisco often tests the misconception that removing a protected attribute (like age) is sufficient to eliminate bias, when in fact proxy features can perpetuate discrimination, making a bias analysis the necessary first step.

How to eliminate wrong answers

Option B is wrong because apologizing and offering a manual review is a reactive customer service response that does not address the root cause of the bias in the model; it fails to assess or remediate the systemic issue. Option C is wrong because immediately removing age from the feature set and retraining the model is a hasty action that could mask the problem without understanding whether age is a proxy for other correlated features (e.g., years of experience), potentially leading to unintended discrimination or model degradation. Option D is wrong because ignoring the complaint outright violates ethical AI principles and regulatory requirements (such as EEOC guidelines), and age is not a legitimate business requirement for resume screening unless it is a bona fide occupational qualification, which is rarely the case.

Full explanation →

825

MCQeasy

An organization deploys an AI model on edge devices for real-time image classification. Which metric is most important to monitor for ensuring the device's operational health?

A.Model calibration error

B.Inference memory consumption

C.Average prediction confidence

D.Model accuracy on local test data

AnswerB

Memory is a key operational health indicator for edge devices.

Why this answer

For edge devices with limited resources, inference memory consumption is the most critical operational health metric because exceeding available memory can cause the model to crash or the device to become unresponsive. Unlike accuracy or confidence, memory usage directly reflects whether the device can sustain real-time inference without resource exhaustion.

Exam trap

CompTIA often tests the misconception that model accuracy or confidence is the primary concern for operational health, but the trap here is that edge device stability depends on resource constraints like memory, not model performance metrics.

How to eliminate wrong answers

Option A is wrong because model calibration error measures the reliability of predicted probabilities, not the operational health of the device. Option C is wrong because average prediction confidence indicates model certainty, not whether the device has sufficient memory to run inference. Option D is wrong because model accuracy on local test data evaluates model performance, not the device's ability to operate without memory overflow or system failure.

Full explanation →

CompTIA AI+ AI0-001 (AI0-001) — Questions 751–825