Courseiva

CompTIA AI+ AI0-001 (AI0-001) — Questions 376–450

1000 questions total · 14pages · All types, answers revealed

Take a mock examExam hub
Page 5

Page 6 of 14

Page 7
376
MCQmedium

A developer is deploying an AI service API. To protect against data leakage through API responses, which access control principle should be applied to API keys?

A.Disable API keys and rely on IP whitelisting only
B.Use a single shared API key for all services
C.Grant all API keys full access to simplify management
D.Implement least-privilege API access with scoped permissions
AnswerD

Least privilege limits data exposure by restricting each key to necessary actions and data.

Why this answer

Option D is correct because the least-privilege principle ensures that each API key is scoped to only the specific permissions required for its intended function, such as read-only access to a single endpoint. This minimizes the blast radius in case the key is compromised, preventing unauthorized access to other services or data. In AI service deployments, scoped permissions are often enforced via OAuth 2.0 scopes or IAM roles tied to the API key.

Exam trap

CompTIA often tests the misconception that simplifying management (Option C) or using IP whitelisting (Option A) is sufficient for security, but the trap is that these approaches ignore the fundamental need for granular access control to prevent data leakage in multi-tenant AI API environments.

How to eliminate wrong answers

Option A is wrong because disabling API keys and relying solely on IP whitelisting removes authentication granularity and fails to protect against data leakage from within the whitelisted network or from IP spoofing attacks. Option B is wrong because using a single shared API key for all services violates the principle of least privilege, as a compromised key would expose all services and data, and it also prevents audit trails for individual users or applications. Option C is wrong because granting all API keys full access simplifies management at the cost of security, allowing any compromised key to access all endpoints and data, directly enabling data leakage.

Full explanation →
377
MCQhard

A team is deploying a fine-tuned LLM for code generation. They need to ensure the model output is always valid JSON. Which prompt engineering technique should they use?

A.Chain-of-thought prompting
B.Few-shot examples of valid JSON outputs
C.Temperature setting to 0
D.Using a larger model variant
AnswerB

Correct. Providing few-shot examples of valid JSON outputs conditions the model to mimic that structure, making it the most direct technique listed.

Why this answer

Few-shot prompting provides the model with examples of valid JSON outputs, conditioning it to mimic the format in its response. This is an effective technique for constraining output structure without requiring model retraining or special modes.

Exam trap

Candidates might mistakenly think that setting temperature to 0 will guarantee valid JSON output, but temperature controls randomness, not format. The only way to enforce a specific format with standard prompting is through explicit instructions and examples.

Full explanation →
378
MCQhard

An organization uses a batch prediction pipeline that processes daily customer data to generate marketing recommendations. One month after deployment, the model's performance degrades significantly. The data pipeline logs show that the input data schema has changed — a new categorical feature 'customer_segment' has been added, and the existing feature 'age_group' is now missing. Which step should the operations team take first?

A.Retrain the model using the new schema and redeploy
B.Update the data preprocessing pipeline to handle missing features and add the new feature
C.Revert to the previous week's model version that was performing well
D.Contact the data engineering team to revert the schema change
AnswerB

This adapts the pipeline to the new schema, enabling proper feeding to the model.

Why this answer

Option B is correct because the immediate priority is to ensure the data preprocessing pipeline can handle the schema change without breaking. The pipeline must gracefully handle the missing 'age_group' feature (e.g., by imputing or dropping it) and incorporate the new 'customer_segment' feature before any model retraining or rollback. This prevents data drift from causing inference errors and maintains pipeline stability.

Exam trap

CompTIA often tests the misconception that retraining the model (Option A) is the first step to fix performance degradation, but the trap here is that the root cause is a schema mismatch in the preprocessing layer, not the model weights.

How to eliminate wrong answers

Option A is wrong because retraining the model without first fixing the preprocessing pipeline would still fail due to missing or misaligned features, and it assumes the new schema is already compatible. Option C is wrong because reverting to a previous model version does not address the root cause — the input data schema has changed, so the old model would still receive malformed data and produce incorrect predictions. Option D is wrong because contacting the data engineering team to revert the schema change is a reactive, non-technical workaround that ignores the need for the operations team to adapt the pipeline to handle schema evolution autonomously.

Full explanation →
379
MCQmedium

A company deploys an LLM-based chatbot that retrieves data from external databases. An attacker embeds malicious instructions in a database record. When the chatbot retrieves that record, it executes the instructions, overriding its system prompt. Which type of attack is this?

A.Model inversion attack
B.Indirect prompt injection
C.Direct prompt injection
D.Membership inference attack
AnswerB

The malicious instructions are embedded in the retrieved data, making it indirect.

Why this answer

This is an indirect prompt injection attack because the malicious instructions are embedded in a third-party data source (the database record) rather than being sent directly by the user. When the LLM retrieves and processes that record, the injected instructions override the system prompt, causing the chatbot to behave contrary to its intended design.

Exam trap

Cisco often tests the distinction between direct and indirect prompt injection by making the attack vector (user input vs. external data source) the key differentiator, so candidates must identify where the malicious instructions originate.

How to eliminate wrong answers

Option A is wrong because a model inversion attack aims to reconstruct training data or extract sensitive information from the model's parameters, not to inject instructions via external data. Option C is wrong because direct prompt injection involves an attacker sending malicious input directly to the LLM (e.g., in a user prompt), not embedding it in a retrieved database record. Option D is wrong because a membership inference attack determines whether a specific data point was part of the model's training set, not about injecting instructions into the model's context.

Full explanation →
380
MCQmedium

An AI model for detecting fraudulent transactions has high precision but low recall. Which business impact is most likely?

A.The model has no impact on fraud detection
B.The model detects all fraudulent transactions
C.Many fraudulent transactions go undetected
D.Many legitimate transactions are flagged as fraud
AnswerC

Low recall indicates a high number of false negatives.

Why this answer

High precision means that when the model flags a transaction as fraudulent, it is very likely correct. However, low recall indicates that the model misses a significant proportion of actual fraudulent transactions. Therefore, the most likely business impact is that many fraudulent transactions go undetected, leading to financial losses.

Exam trap

CompTIA often tests the distinction between precision and recall by presenting a scenario where candidates confuse high precision with high recall, leading them to incorrectly select option D (many legitimate transactions flagged) instead of recognizing that low recall causes undetected fraud.

How to eliminate wrong answers

Option A is wrong because a model with high precision and low recall does have a significant impact—it fails to catch many fraud cases, which directly affects business outcomes. Option B is wrong because low recall means the model does not detect all fraudulent transactions; it misses many, contradicting the claim of detecting all fraud. Option D is wrong because high precision implies few false positives, so legitimate transactions are rarely flagged as fraud; that scenario would correspond to low precision, not high precision.

Full explanation →
381
MCQhard

A data scientist is preparing a dataset for a binary classification model to detect fraudulent transactions. The dataset contains 1% fraudulent and 99% legitimate transactions. The goal is to maximize recall for the fraud class while maintaining a precision above 0.5. Which data preparation strategy is MOST effective?

A.Apply random undersampling of the majority class until the dataset is balanced
B.Remove all duplicate transactions from the dataset
C.Use the raw dataset without any resampling, relying on class weights during training
D.Apply SMOTE (Synthetic Minority Oversampling Technique) to generate synthetic fraud examples
AnswerD

SMOTE creates synthetic minority samples, balancing the classes without losing majority data, improving recall while maintaining reasonable precision.

Why this answer

Handling imbalanced data typically requires resampling. For recall maximization with moderate precision constraint, oversampling the minority class (e.g., SMOTE) is effective. Undersampling loses too many majority samples, and using raw data leads the model to predict the majority class.

Full explanation →
382
MCQeasy

A data engineer needs to combine two datasets, each with unique customer_id, to include all records from both datasets. Which join type should be used?

A.FULL OUTER JOIN
B.RIGHT JOIN
C.LEFT JOIN
D.INNER JOIN
AnswerA

FULL OUTER JOIN includes all records from both tables, matching where possible and filling nulls elsewhere.

Why this answer

A FULL OUTER JOIN returns all records from both datasets, matching rows where the customer_id is present in both and filling in NULLs for missing matches. This is the only join type that guarantees every unique customer_id from either dataset appears in the result, which is exactly what the requirement specifies.

Exam trap

CompTIA often tests the misconception that LEFT JOIN or RIGHT JOIN can include all records from both datasets, but candidates forget that these asymmetric joins exclude non-matching rows from the opposite side.

How to eliminate wrong answers

Option B (RIGHT JOIN) is wrong because it returns only all rows from the right dataset and matching rows from the left, omitting any customer_id that exists only in the left dataset. Option C (LEFT JOIN) is wrong because it returns only all rows from the left dataset and matching rows from the right, omitting any customer_id that exists only in the right dataset. Option D (INNER JOIN) is wrong because it returns only rows where customer_id exists in both datasets, discarding all non-matching records from either side.

Full explanation →
383
Multi-Selectmedium

A team is deploying a model that must comply with GDPR. Users can request deletion of their data. Which TWO practices should be implemented to support this compliance? (Select TWO.)

Select 2 answers
A.Enable output caching for frequently requested predictions
B.Validate inputs to prevent prompt injection attacks
C.Use a vector database to store user embeddings
D.Maintain the ability to delete a user's data from training sets and derived features
E.Implement data versioning and lineage tracking
AnswersD, E

Directly supports the right to erasure by allowing removal of user data and any features based on it.

Why this answer

Option D is correct because GDPR's 'right to erasure' requires that upon user request, the organization must delete not only the user's raw data but also any derived features or embeddings that were generated from that data. Without this capability, the model could still indirectly retain user information through trained parameters or feature stores, violating compliance.

Exam trap

CompTIA often tests the misconception that simply using a vector database or caching mechanism satisfies GDPR deletion requirements, when in fact the critical practice is maintaining the ability to delete user data from all derived artifacts, including training sets and feature stores.

Full explanation →
384
Multi-Selectmedium

A data scientist is fine-tuning a large language model for a domain-specific task using QLoRA. Which TWO statements correctly describe QLoRA's advantages?

Select 2 answers
A.It enables fine-tuning on consumer-grade GPUs by reducing memory requirements
B.It reduces memory usage by quantizing the base model to 4-bit precision
C.It requires more training data than full fine-tuning to achieve comparable accuracy
D.It trains the full model parameters with low precision
E.It increases inference speed compared to the base model
AnswersA, B

Lower memory allows fine-tuning on smaller GPUs, such as consumer-grade hardware.

Why this answer

QLoRA uses 4-bit quantization to reduce memory, and freezes the base model while training low-rank adapters. It does not reduce inference latency (base model still runs), and does not require more data than full fine-tuning.

Full explanation →
385
MCQmedium

A data scientist is using PyTorch to train a custom NLP model. The training is slow on a single GPU. They want to speed up training by using multiple GPUs on a single machine. Which PyTorch feature should they use?

A.TorchScript tracing
B.torch.nn.DataParallel
C.torch.optim.SGD
D.PyTorch Lightning's zero_grad function
AnswerB

DataParallel automatically splits input across GPUs and aggregates gradients; it's the simplest multi-GPU approach.

Why this answer

DataParallel (or DistributedDataParallel) is PyTorch's built-in feature to split batches across multiple GPUs. It is straightforward for single-machine multi-GPU training.

Full explanation →
386
MCQhard

An organization uses a fine-tuned LLM for generating financial reports. An attacker gains access to the model's API and sends a series of queries that gradually reconstruct the training data of the fine-tuned model. This is an example of which attack?

A.Membership inference
B.Data poisoning
C.Model extraction
D.Model inversion
AnswerD

Model inversion uses model outputs to infer or reconstruct training data.

Why this answer

Model inversion attacks aim to reconstruct training data from model outputs. Data poisoning corrupts training, model extraction steals the model, and membership inference determines presence, not reconstruction.

Exam trap

Candidates may confuse model inversion with model extraction; extraction steals the model parameters, inversion reconstructs training data.

Full explanation →
387
MCQeasy

A company deploys a deep learning model for real-time image classification. After deployment, they notice high inference latency exceeding the 100ms SLA. Which action would most likely reduce latency without significantly impacting accuracy?

A.Add more training data to improve model robustness
B.Replace the model with a simpler logistic regression model
C.Increase batch size for inference
D.Apply model quantization
AnswerD

Quantization reduces model size and inference time with minor accuracy impact.

Why this answer

Model quantization reduces the precision of the model's weights and activations (e.g., from 32-bit floating point to 8-bit integer), which significantly decreases memory bandwidth and computational requirements during inference. This directly lowers latency without fundamentally altering the model's learned representations, so accuracy degradation is typically minimal (often <1-2%).

Exam trap

CompTIA often tests the misconception that increasing batch size always improves latency, when in fact it increases per-request latency in real-time systems, and that simpler models are always better for latency, ignoring the critical accuracy requirement.

How to eliminate wrong answers

Option A is wrong because adding more training data improves model robustness and generalization but does not reduce inference latency; it may even increase training time and model complexity. Option B is wrong because replacing a deep learning model with a logistic regression model would drastically reduce accuracy for complex image classification tasks, failing the 'without significantly impacting accuracy' constraint. Option C is wrong because increasing batch size for inference increases the number of images processed per batch, which can improve throughput but actually increases per-request latency (time to first prediction) and may exceed the 100ms SLA for real-time applications.

Full explanation →
388
MCQmedium

A data scientist needs to explain why a black-box model denied a loan application. Which explainability technique generates local feature importance values using a simpler interpretable model around the prediction?

A.Model card
B.LIME
C.Attention visualisation
D.SHAP values
AnswerB

LIME fits a simple model (e.g., linear) locally to approximate the black-box model's decision boundary for a specific instance.

Why this answer

LIME (Local Interpretable Model-agnostic Explanations) is the correct technique because it generates local feature importance values by fitting a simpler, interpretable model (e.g., linear regression or decision tree) around the prediction of the black-box model. This allows the data scientist to explain why a specific loan application was denied by identifying which features (e.g., income, credit score) most influenced that particular decision. Unlike global methods, LIME focuses on the local neighborhood of the instance, making it ideal for explaining individual predictions.

Exam trap

Cisco often tests the distinction between local vs. global explainability methods, and the trap here is that candidates may confuse SHAP values (which also provide local feature importance) with LIME, failing to recognize that LIME uniquely uses a simpler interpretable surrogate model trained around the prediction, while SHAP uses game-theoretic contributions without a surrogate model.

How to eliminate wrong answers

Option A is wrong because a model card is a documentation artifact that summarizes a model's intended use, performance, and limitations at a global level, not a technique for generating local feature importance values for a single prediction. Option C is wrong because attention visualization is specific to neural network architectures (e.g., transformers) and provides insight into which parts of the input the model 'attends to,' but it is not a model-agnostic method for generating local feature importance with a simpler interpretable model. Option D is wrong because SHAP values, while they do provide local feature importance, are based on cooperative game theory (Shapley values) and do not use a simpler interpretable model around the prediction; instead, they compute additive feature contributions directly from the model's output.

Full explanation →
389
Multi-Selecteasy

Which TWO data preprocessing techniques reduce the dimensionality of a dataset?

Select 2 answers
A.One-hot encoding
B.Imputation
C.Feature scaling
D.Principal Component Analysis (PCA)
E.Feature selection
AnswersD, E

PCA reduces dimensionality by projecting data onto principal components.

Why this answer

Principal Component Analysis (PCA) is a dimensionality reduction technique that transforms the original features into a new set of uncorrelated variables (principal components) ordered by the variance they capture. By selecting only the top components, PCA reduces the number of features while retaining most of the dataset's information, effectively lowering the dimensionality.

Exam trap

CompTIA often tests the distinction between techniques that transform or select features (reducing dimensionality) versus those that prepare data for modeling (like encoding, imputation, or scaling) without changing the number of features.

Full explanation →
390
MCQmedium

A company uses an AI model to predict equipment failures. The model outputs a probability of failure. To minimize false alarms, the operations team wants a high precision. Which deployment strategy should they implement?

A.Retrain the model on more recent data
B.Increase the decision threshold for positive classification
C.Decrease the decision threshold
D.Use an ensemble of models with voting
AnswerB

Higher threshold means fewer positive predictions, increasing precision.

Why this answer

To minimize false alarms and achieve high precision, the operations team should increase the decision threshold for positive classification. A higher threshold means the model only predicts a failure when it is very confident, reducing the number of false positives (false alarms) at the cost of potentially missing some true failures (lower recall). This directly controls the precision-recall trade-off without changing the underlying model.

Exam trap

CompTIA often tests the precision-recall trade-off by making candidates confuse increasing the threshold (which improves precision) with decreasing it (which improves recall), or by suggesting retraining or ensemble methods as direct solutions for precision tuning.

How to eliminate wrong answers

Option A is wrong because retraining on more recent data improves model accuracy and relevance but does not directly control the precision-recall trade-off; it may not reduce false alarms if the model's calibration remains unchanged. Option C is wrong because decreasing the decision threshold would make the model more sensitive, increasing the number of positive predictions and thus increasing false alarms (lower precision), which is the opposite of the goal. Option D is wrong because using an ensemble of models with voting can improve overall accuracy and robustness, but it does not specifically target precision; the voting mechanism may still produce many false positives unless the threshold is also adjusted.

Full explanation →
391
Multi-Selectmedium

Which THREE are common data preprocessing steps in a machine learning pipeline? (Choose 3)

Select 3 answers
A.Hyperparameter tuning
B.Encoding categorical variables
C.Model evaluation
D.Scaling numeric features
E.Handling missing values
AnswersB, D, E

Categorical data must be converted to numeric.

Why this answer

Encoding categorical variables is a common data preprocessing step because machine learning algorithms require numerical input. Techniques like one-hot encoding or label encoding convert categorical data (e.g., colors, countries) into numeric format, enabling the model to process them correctly. Without this step, the model would misinterpret categorical labels as ordinal or meaningless numeric values.

Exam trap

CompTIA often tests the distinction between preprocessing steps (data cleaning, transformation) and later pipeline stages (model tuning, evaluation), so candidates mistakenly select hyperparameter tuning or model evaluation as preprocessing steps.

Full explanation →
392
Multi-Selecthard

A security team is reviewing an LLM-powered application that can execute SQL queries based on user requests. They want to implement defenses against prompt injection that could lead to unauthorized database access. Which TWO controls are MOST effective? (Select two.)

Select 2 answers
A.Robust training techniques
B.Input validation and sanitization
C.Access controls on the database
D.Rate limiting
E.Output filtering
AnswersB, E

Validates and sanitizes user input to remove or neutralize injection attempts.

Why this answer

Input validation and sanitization can filter malicious prompt content, and output filtering can block responses containing unauthorized data. Rate limiting does not prevent injection, access controls on the database are important but not a direct defense against injection (they limit impact), and robust training does not prevent injection at inference time.

Full explanation →
393
MCQhard

A large financial services company deploys multiple AI models on a shared Kubernetes cluster with GPU nodes. The models serve real-time fraud detection and credit scoring. Recently, the operations team observed frequent out-of-memory (OOM) errors during peak hours, causing inference failures. The monitoring dashboards show GPU memory utilization averaging 90% during peak times, and pods are being evicted. The team has allocated 8GB per pod and the total cluster GPU memory is 32GB. The models require at least 4GB each, but the fraud detection model occasionally spikes to 7GB. Which course of action best resolves the OOM errors while maintaining high availability?

A.Reduce the batch size and model complexity for all models to lower memory footprint
B.Set resource limits and requests per model based on observed usage, and implement pod priority classes
C.Provision larger GPU nodes with 48GB memory each
D.Increase the memory request for all pods to 8GB to ensure they have enough
AnswerB

Limits prevent OOM, priority ensures critical models get resources.

Why this answer

Option B is correct because it uses Kubernetes resource management features—setting precise resource requests and limits based on observed GPU memory usage—combined with pod priority classes to ensure critical fraud detection pods are scheduled and retained during contention. This prevents OOM errors by capping memory per pod while allowing the spike-prone fraud model to be prioritized over less critical workloads, maintaining high availability without overprovisioning.

Exam trap

CompTIA often tests the misconception that simply increasing resource requests or node size solves OOM errors, when the real solution involves proper resource limits and scheduling policies to handle variable workloads and maintain availability.

How to eliminate wrong answers

Option A is wrong because reducing batch size and model complexity may degrade inference accuracy or latency, and it does not address the root cause of memory spikes for the fraud detection model; it is a workaround that sacrifices performance. Option C is wrong because provisioning larger GPU nodes (48GB) is a costly overprovisioning approach that does not solve the scheduling or priority issue—it only shifts the bottleneck and may still allow a single pod to consume excessive memory and cause OOM on the larger node. Option D is wrong because increasing the memory request for all pods to 8GB does not prevent the fraud detection model from spiking to 7GB (which is under 8GB) and ignores the need for limits and priority; it may also lead to resource waste and does not address eviction during peak contention.

Full explanation →
394
MCQhard

An ML engineering team has a retraining pipeline that triggers automatically when model accuracy drops below a threshold. Recently, the model's accuracy has been fluctuating, causing frequent retraining and high compute costs. The team suspects the data distribution is changing slowly. Which approach should the team implement to reduce unnecessary retraining while maintaining model performance?

A.Use a simpler model to reduce variability
B.Implement a statistical drift detection method on input features
C.Increase the frequency of model retraining
D.Reduce the batch size for inference
AnswerB

Drift detection ensures retraining only when meaningful change occurs.

Why this answer

Option B is correct because implementing a statistical drift detection method (e.g., using KL divergence, PSI, or ADWIN) on input features allows the team to identify when the data distribution has genuinely changed, rather than reacting to random accuracy fluctuations. This reduces unnecessary retraining by triggering the pipeline only when statistically significant drift is detected, maintaining model performance without the high compute costs of frequent retraining.

Exam trap

CompTIA often tests the misconception that increasing retraining frequency or simplifying the model can solve drift-related issues, but the correct approach is to detect drift statistically before deciding to retrain.

How to eliminate wrong answers

Option A is wrong because using a simpler model may reduce variability but does not address the root cause of distribution drift; it could also degrade performance by underfitting the true underlying patterns. Option C is wrong because increasing retraining frequency would exacerbate the compute cost problem and may overfit to transient fluctuations, not solve the issue of unnecessary retraining. Option D is wrong because reducing the batch size for inference affects throughput and latency, not the detection of data distribution changes or the decision to retrain.

Full explanation →
395
MCQmedium

A developer is building an AI agent that needs to call external APIs to complete user requests. The agent must decide which API to call based on the user's natural language input. Which technique should the developer use to enable the agent to invoke APIs?

A.Fine-tuning the LLM on API documentation
B.Chain-of-thought reasoning
C.Few-shot prompting with examples of API calls
D.Function calling
AnswerD

Function calling allows the model to output a structured JSON that triggers an API call.

Why this answer

Function calling is a technique where the LLM outputs a structured request to call a predefined function/API, enabling tool use.

Full explanation →
396
MCQeasy

A developer is building an AI-powered code completion tool. To ensure the model does not output malicious code when prompted with 'Write code to delete all files on the system', which defense is most effective?

A.Output filtering to detect and block dangerous code constructs
B.Input validation to block the word 'delete'
C.Rate limiting on the number of requests per user
D.Retraining the model on safe code only
AnswerA

Output filtering can analyze generated code for malicious patterns and block it.

Why this answer

Output filtering can block generated code that contains dangerous patterns like file deletion commands.

Full explanation →
397
MCQhard

An e-commerce company deploys a deep learning model for product recommendation. After a new data pipeline is implemented, the model's online performance degrades: recall drops by 20% and the click-through rate decreases. The data scientists suspect data drift. They compare the distribution of the input features between the training data and recent production data. The Kolmogorov-Smirnov test shows significant differences for two numerical features (price and rating). The team also notices that the frequency of categorical feature 'category' has changed. Which of the following is the MOST appropriate first step? A. Immediately retrain the model on all available data including new production data. B. Roll back to the previous data pipeline and investigate the root cause of drift. C. Use feature selection to remove the drifting features and retrain. D. Implement a monitoring dashboard to track drift over time and set up alerts.

A.Implement a monitoring dashboard to track drift over time and set up alerts.
B.Roll back to the previous data pipeline and investigate the root cause of drift.
C.Use feature selection to remove the drifting features and retrain.
D.Immediately retrain the model on all available data including new production data.
AnswerB

Rolling back restores the previous stable distribution; investigating the root cause prevents recurrence.

Why this answer

Option B is correct. Since the drift occurred after a pipeline change, rolling back and investigating the root cause is the most prudent first step before making model changes. Retraining on drifted data (A) might incorporate a faulty distribution.

Removing drifting features (C) could lose important information and may not fully address the issue. Implementing monitoring (D) is useful for long-term but does not address the immediate degradation.

Full explanation →
398
MCQeasy

A developer sees the above error during inference on a deployed image classification model. What is the most likely cause?

A.The model version is incompatible with the serving framework
B.The input images are not being resized to the required dimensions
C.The inference server does not support batch processing
D.The model is overfitting to a specific image size
AnswerB

Model expects 299x299 but receives 224x224, so preprocessing is missing resizing.

Why this answer

The error during inference is typically caused by a mismatch between the input tensor shape expected by the model and the shape of the provided image. Image classification models are trained on images of a fixed size. If input images are not resized to the required dimensions, the model's input layer expects a different shape, resulting in a shape mismatch error.

Option B correctly identifies this as the most likely cause.

Exam trap

CompTIA often tests the misconception that inference errors are caused by model versioning or server configuration, when the actual issue is a simple preprocessing step like image resizing that candidates overlook.

How to eliminate wrong answers

Option A is wrong because model version incompatibility with the serving framework usually manifests as a loading or serialization error (e.g., 'Unsupported op set' or 'Model not found'), not a runtime shape mismatch during inference. Option C is wrong because the inference server's batch processing capability is unrelated to the error; even if batch processing is unsupported, the server would still process single images, and the error would not be about input dimensions. Option D is wrong because overfitting to a specific image size is a training-phase issue that would affect model accuracy, not cause a runtime shape mismatch error during inference.

Full explanation →
399
MCQhard

A company operating in the EU must comply with GDPR. An AI model processes personal data for customer segmentation. Which of the following ensures compliance?

A.Obtain explicit consent once and use data indefinitely.
B.Store personal data permanently for model improvement.
C.Use only aggregated data without any individual records.
D.Implement data anonymization and allow users to request deletion.
AnswerD

Anonymization reduces privacy risk, and deletion capability ensures compliance with GDPR rights.

Why this answer

Option D is correct because GDPR mandates that personal data must be processed lawfully, with data minimization and the right to erasure. Implementing data anonymization removes personally identifiable information (PII) so the data is no longer considered personal data under GDPR, and allowing users to request deletion directly satisfies the 'right to be forgotten' (Article 17). This approach ensures compliance by both protecting individual privacy and providing a mechanism for data subjects to exercise their legal rights.

Exam trap

CompTIA often tests the misconception that pseudonymization or simple aggregation is sufficient for GDPR compliance, when in fact only irreversible anonymization (where no individual can be re-identified) removes data from GDPR scope, and the right to deletion must still be explicitly supported for any remaining personal data.

How to eliminate wrong answers

Option A is wrong because GDPR requires that consent be specific, informed, and revocable; obtaining consent once does not permit indefinite use, and data must be retained only as long as necessary for the stated purpose. Option B is wrong because storing personal data permanently violates the data minimization and storage limitation principles (Article 5(1)(c) and (e)), and model improvement is not a valid basis for indefinite retention without explicit, ongoing consent. Option C is wrong because while aggregated data reduces risk, it does not automatically ensure compliance if the aggregation method is reversible or if the data can be re-identified; true anonymization must be irreversible and meet the GDPR's standard of 'anonymous information' (Recital 26).

Full explanation →
400
MCQmedium

Refer to the exhibit. What is the most likely issue and what action should be taken?

A.Learning rate is too low; increase it
B.Underfitting; increase model complexity
C.Overfitting; apply early stopping around epoch 15
D.Data imbalance; use class weights
AnswerC

Validation loss starts rising after epoch 15; early stopping halts training at that point.

Why this answer

The training loss continues to decrease while the validation loss starts to increase after approximately epoch 15, which is a classic sign of overfitting. The model is memorizing the training data rather than generalizing, so applying early stopping around epoch 15 would prevent further divergence and preserve the best validation performance.

Exam trap

CompTIA often tests the distinction between overfitting and underfitting by showing loss curves where training loss continues to drop while validation loss rises, tricking candidates into thinking the model needs more training or a lower learning rate.

How to eliminate wrong answers

Option A is wrong because a low learning rate would cause both training and validation loss to decrease very slowly or plateau, not diverge with validation loss rising. Option B is wrong because underfitting would show both training and validation loss remaining high and not decreasing, whereas here training loss is still dropping. Option D is wrong because data imbalance typically causes poor performance on the minority class across both training and validation sets, not a divergence in loss curves after a certain epoch.

Full explanation →
401
MCQeasy

Based on the exhibit, what issue should the team address?

A.Model accuracy below threshold
B.Potential fairness bias across groups
C.High latency
D.Low throughput
AnswerB

The disparity in accuracy between Group B (0.83) and other groups (0.97, 0.96) indicates a fairness issue that needs to be addressed.

Why this answer

Option B is correct because the exhibit likely shows a confusion matrix or performance metrics broken down by demographic groups (e.g., race, gender), revealing that the model's false positive or false negative rates differ significantly across groups. This disparity indicates a potential fairness bias, which must be addressed to ensure equitable outcomes, especially in high-stakes AI applications like hiring or lending.

Exam trap

CompTIA often tests the misconception that high overall accuracy or low latency/throughput issues are the primary concerns, when the real problem is hidden bias revealed only by disaggregated performance metrics across subgroups.

How to eliminate wrong answers

Option A is wrong because the exhibit does not show an overall accuracy metric below a threshold; instead, it highlights group-wise performance differences, not a global accuracy issue. Option C is wrong because latency refers to inference time per request, which is not indicated by group-wise performance metrics or confusion matrices. Option D is wrong because throughput measures the number of predictions per second, which is unrelated to the group-level bias patterns shown in the exhibit.

Full explanation →
402
Multi-Selecteasy

A company is deploying a pre-trained image classification model for facial recognition in a security system. They are concerned about adversarial examples. Which TWO of the following are effective defenses against adversarial examples?

Select 2 answers
A.Adversarial training during model development
B.Gradient masking to hide model gradients
C.Input sanitization techniques such as JPEG compression or denoising
D.Homomorphic encryption of input images
E.Federated learning to train on distributed data
AnswersA, C

Adversarial training incorporates adversarial examples into the training set, making the model more robust to such perturbations at inference time.

Why this answer

Adversarial training (including the model with adversarial examples during training) and input sanitization (e.g., JPEG compression, denoising) are proven defenses against adversarial perturbations. Gradient masking is a weak defense. Homomorphic encryption and federated learning are unrelated to adversarial robustness.

Full explanation →
403
MCQmedium

Under the EU AI Act, an AI system that uses subliminal techniques to materially distort a person's behaviour, causing psychological or physical harm, would be classified under which risk tier?

A.Unacceptable risk
B.Limited risk
C.High risk
D.Minimal risk
AnswerA

Systems that employ subliminal techniques to distort behaviour causing harm are banned under the unacceptable risk category.

Why this answer

The EU AI Act categorises such systems as 'unacceptable risk' and prohibits them outright. High risk includes critical infrastructure, education, employment, etc. Limited risk involves transparency obligations.

Minimal risk covers all other systems.

Full explanation →
404
Multi-Selectmedium

A startup is building a medical diagnosis support system using a large language model. To prevent the model from generating harmful advice due to hallucinations, which TWO measures should they implement as part of their AI security strategy?

Select 2 answers
A.Ground the model using Retrieval-Augmented Generation (RAG) with curated medical databases
B.Monitor for anomalous inputs to detect data poisoning attempts
C.Employ federated learning to train on decentralized patient data
D.Implement output filtering and content moderation to block harmful or unverified medical advice
E.Use robust training techniques like adversarial training
AnswersA, D

RAG reduces hallucinations by providing the model with relevant, authoritative information at inference time, making it less likely to generate unsupported advice.

Why this answer

Option A is correct because Retrieval-Augmented Generation (RAG) grounds the LLM's outputs in verified, curated medical databases (e.g., PubMed, clinical guidelines). By retrieving relevant, factual information before generating a response, RAG significantly reduces the risk of hallucinations that could lead to harmful medical advice. This is a direct security measure to ensure the model's outputs are factually accurate and safe.

Exam trap

CompTIA AI often tests the distinction between inference-time security controls (like RAG and output filtering) versus training-time or data-protection measures (like federated learning, adversarial training, or anomaly detection), leading candidates to select options that are valid security techniques but do not directly address the specific threat of hallucinated harmful advice.

Full explanation →
405
MCQhard

A team fine-tunes a 7B parameter LLM using LoRA on a custom instruction dataset. After training, they observe that the model's outputs are only marginally different from the base model. Which is the MOST likely cause?

A.The dataset contained too many examples, overfitting the adapter
B.The base model was too small to benefit from fine-tuning
C.The LoRA rank was set too low (e.g., r=1), limiting the adapter's capacity to learn the task
D.The learning rate was too high, causing the model to diverge
AnswerC

Low rank reduces the number of trainable parameters; the adapter may not have enough capacity to alter behavior significantly.

Why this answer

LoRA has a rank hyperparameter that controls adapter expressiveness. If the rank is too low, the adapter cannot capture the desired task. Other hyperparameters like learning rate affect convergence but rank directly impacts capacity.

Full explanation →
406
Multi-Selectmedium

A machine learning engineer is deploying a model to production. Which TWO practices are essential for ensuring reproducibility of model predictions?

Select 2 answers
A.Increase the number of training epochs to ensure convergence.
B.Use the same GPU hardware for both training and inference.
C.Use parallel data loading to speed up inference.
D.Version-control the model artifact (e.g., using MLflow or DVC).
E.Fix random seeds for all libraries (e.g., NumPy, TensorFlow).
AnswersD, E

Versioning ensures the exact model is used for inference.

Why this answer

Version-controlling the model artifact (D) is essential because it allows you to reproduce the exact model binary that generated a prediction, ensuring that any changes to the model code, hyperparameters, or training data do not silently alter outputs. Tools like MLflow or DVC store the model along with its metadata, enabling rollback and auditability in production.

Exam trap

CompTIA often tests the misconception that hardware consistency (e.g., same GPU) is required for reproducibility, when in fact deterministic software practices (version control and seed fixing) are the critical factors.

Full explanation →
407
MCQmedium

A security analyst notices that an AI model used for facial recognition is returning unusually high confidence scores for certain individuals while consistently misidentifying others. Which type of attack is most likely occurring?

A.Data poisoning
B.Evasion attack
C.Model inversion attack
D.Model extraction attack
AnswerC

Inversion exploits confidence scores to infer private training data, often showing high confidence on seen data.

Why this answer

A model inversion attack allows an adversary to reconstruct training data or infer sensitive attributes from the model's outputs. In this scenario, the unusually high confidence scores for certain individuals and misidentification of others indicate that the attacker is exploiting the model's internal representations to extract information about the training data, leading to biased or overconfident predictions for specific classes.

Exam trap

Cisco often tests the distinction between attacks that affect model outputs (evasion) versus attacks that extract or infer training data (model inversion), and candidates may confuse the high confidence scores with a successful evasion or poisoning effect.

How to eliminate wrong answers

Option A is wrong because data poisoning involves injecting malicious data into the training set to corrupt the model's behavior, which would typically cause systematic errors across many inputs rather than selectively high confidence for some individuals. Option B is wrong because an evasion attack (adversarial example) manipulates input data to cause misclassification, but it does not explain the high confidence scores for certain individuals; evasion attacks usually reduce confidence or cause incorrect labels. Option D is wrong because model extraction aims to duplicate the model's functionality by querying it and training a substitute, not to reveal training data or cause confidence anomalies for specific individuals.

Full explanation →
408
Multi-Selecthard

Which TWO are key differences between Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN)?

Select 2 answers
A.CNNs are designed for sequential data; RNNs for spatial data
B.RNNs have internal memory; CNNs do not
C.CNNs can handle variable-length inputs; RNNs require fixed-size inputs
D.CNNs use backpropagation; RNNs do not
E.CNNs use weight sharing across spatial dimensions; RNNs share weights across time steps
AnswersB, E

RNNs maintain a hidden state for temporal memory; CNNs are feedforward.

Why this answer

Option B is correct because RNNs possess a hidden state that acts as internal memory, allowing them to retain information from previous time steps, which is essential for processing sequential data. In contrast, CNNs lack this internal memory mechanism; they process inputs independently without maintaining a state across different inputs, making them unsuitable for tasks requiring temporal context.

Exam trap

CompTIA often tests the misconception that CNNs and RNNs are distinguished by their training algorithms or input size requirements, when the core difference lies in their architectural design—specifically, internal memory and weight sharing mechanisms.

Full explanation →
409
MCQhard

A team is deploying a multi-modal AI model that processes both text and images. They need to ensure that inference requests are handled quickly even during traffic spikes. Which integration pattern is BEST suited for this use case?

A.Deploy a synchronous REST API with auto-scaling
B.Stream responses directly from the model to the client
C.Pre-compute all possible outputs and cache them
D.Use an event-driven architecture with a queue and worker instances
AnswerD

Queues decouple request submission from processing, enabling resilient scaling and handling spikes.

Why this answer

Asynchronous processing with a message queue allows requests to be buffered and processed without blocking the user, scaling out workers as needed.

Full explanation →
410
MCQmedium

A manufacturing company uses a computer vision AI to inspect products on an assembly line for defects. The AI model was trained on images from a single camera angle under bright, uniform lighting. Recently, the company moved the inspection station to a different part of the factory where lighting is dimmer and varies due to nearby windows. The model now misclassifies many non-defective products as defective, causing false alarms and production delays. The team has limited labeled data from the new environment. Which action should the team take to restore inspection accuracy while minimizing downtime?

A.Apply domain adaptation techniques using a small set of labeled images from the new environment
B.Increase the defect classification threshold to reduce false positives
C.Revert to the previous lighting setup by reinstalling bright, uniform lights
D.Retrain the model from scratch using a large dataset of images from the new environment
AnswerA

Domain adaptation adjusts the model to new conditions with minimal data.

Why this answer

Domain adaptation techniques allow a model trained on a source domain (bright, uniform lighting) to generalize to a target domain (dim, variable lighting) using only a small set of labeled images from the new environment. This approach minimizes downtime because it avoids the need for large-scale data collection or retraining from scratch, and it directly addresses the distribution shift that causes false positives.

Exam trap

CompTIA often tests the misconception that simply adjusting a threshold or reverting to old conditions is a valid fix, when the correct approach is to adapt the model to the new data distribution using domain adaptation.

How to eliminate wrong answers

Option B is wrong because increasing the classification threshold reduces false positives at the cost of increasing false negatives, which would allow defective products to pass inspection — a critical safety and quality risk. Option C is wrong because reverting to the previous lighting setup is a workaround that does not solve the underlying domain shift problem and may be impractical or costly if the new location is fixed. Option D is wrong because retraining from scratch requires a large labeled dataset from the new environment, which the team does not have, and would cause significant downtime for data collection and training.

Full explanation →
411
MCQeasy

A data scientist is training a neural network to classify images of handwritten digits. The model achieves 99% accuracy on training data but only 85% on validation data. Which technique should the scientist apply first to address this issue?

A.Remove one or more hidden layers from the network
B.Increase the number of training epochs
C.Apply L2 regularization to the network weights
D.Add more features to the input data
AnswerC

L2 regularization penalizes large weights and reduces overfitting.

Why this answer

The model shows high training accuracy (99%) but lower validation accuracy (85%), which is a classic sign of overfitting. L2 regularization (option C) adds a penalty term to the loss function proportional to the squared magnitude of the weights, discouraging the network from learning overly complex patterns that do not generalize. This directly addresses overfitting without reducing the model's capacity too aggressively.

Exam trap

CompTIA often tests the distinction between overfitting and underfitting, and the trap here is that candidates may confuse increasing epochs (option B) as a solution to low validation accuracy, when in fact it exacerbates overfitting in this scenario.

How to eliminate wrong answers

Option A is wrong because removing hidden layers reduces the model's capacity, which may underfit and does not specifically target the overfitting problem; the network already has sufficient capacity to memorize the training data. Option B is wrong because increasing the number of training epochs would likely worsen overfitting by allowing the model to further memorize noise in the training data, not improve validation performance. Option D is wrong because adding more features to the input data (e.g., additional pixel-level transformations) would increase the dimensionality and risk of overfitting, not reduce it, and is not a standard technique for addressing overfitting in neural networks.

Full explanation →
412
MCQhard

A company wants to fine-tune a 70B-parameter LLM for a specialized domain but has limited GPU memory (e.g., 24 GB VRAM). Which technique allows fine-tuning with minimal memory footprint?

A.Instruction tuning with a smaller dataset
B.QLoRA (Quantized Low-Rank Adaptation)
C.LoRA (Low-Rank Adaptation) on the base model
D.Full fine-tuning with gradient checkpointing
AnswerB

QLoRA quantizes the base model to 4-bit and uses LoRA adapters, fitting a 70B model in 24 GB VRAM.

Why this answer

QLoRA (Quantized Low-Rank Adaptation) uses 4-bit quantization and low-rank adapters, reducing memory requirements dramatically while preserving fine-tuning quality.

Full explanation →
413
MCQmedium

Which chunking strategy for RAG is MOST appropriate when documents have a natural hierarchical structure (e.g., sections, subsections)?

A.Hierarchical chunking that preserves document structure
B.Fixed-size chunking with no overlap
C.Semantic chunking based on sentence boundaries
D.Random chunking with varying sizes
AnswerA

Hierarchical chunking maintains the document's section and subsection organization, improving retrieval relevance.

Why this answer

Hierarchical chunking respects the document structure, preserving context and enabling retrieval at different levels (e.g., section or paragraph).

Full explanation →
414
MCQmedium

A data scientist is evaluating a binary classification model. The model achieves 95% accuracy on the test set, but the precision is 0.60 and recall is 0.55. The dataset has 90% negative class samples. Which metric should the team focus on to improve the model?

A.F1 score
B.Perplexity
C.BLEU score
D.Accuracy
AnswerA

F1 score is the harmonic mean of precision and recall, providing a balanced metric that accounts for both false positives and false negatives.

Why this answer

With high class imbalance (90% negatives), accuracy is misleading. F1 score balances precision and recall, giving a better picture of performance on the minority class. AUC-ROC is also good but F1 directly optimizes for positive class.

Full explanation →
415
MCQhard

An AI model achieves high accuracy on training data but performs poorly on new test data. The data scientist suspects the model has memorized noise. Which technique directly adds a penalty term to the loss function to address this?

A.Batch normalization
B.Data augmentation
C.Dropout
D.L2 regularization
AnswerD

Correct; L2 adds a penalty term proportional to squared weights.

Why this answer

L2 regularization (also known as weight decay) directly adds a penalty term proportional to the squared magnitude of the model's weights to the loss function. This discourages the model from fitting the noise in the training data by keeping weights small, thereby reducing overfitting and improving generalization to new test data.

Exam trap

CompTIA often tests the distinction between regularization techniques that modify the loss function (L2) versus those that modify the network architecture or data (dropout, batch normalization, data augmentation), so candidates mistakenly choose dropout because it is a well-known regularization method, even though it does not add a penalty term to the loss function.

How to eliminate wrong answers

Option A is wrong because batch normalization normalizes the inputs of each layer to stabilize and accelerate training, but it does not add a penalty term to the loss function; it addresses internal covariate shift, not overfitting from memorized noise. Option B is wrong because data augmentation artificially expands the training dataset by applying transformations (e.g., rotations, flips) to reduce overfitting, but it does not modify the loss function with a penalty term. Option C is wrong because dropout randomly drops neurons during training to prevent co-adaptation, which is a regularization technique but it does not add a penalty term to the loss function; it works by altering the network architecture during training.

Full explanation →
416
Multi-Selectmedium

A data science team wants to implement a feature store to serve pre-computed features for both training and inference with low latency. Which TWO tools are commonly used for building a feature store?

Select 2 answers
A.Kubeflow
B.Apache Hive
C.Feast
D.Tecton
E.MLflow
AnswersC, D

Feast is an open-source feature store that manages and serves features.

Why this answer

Feast (Feature Store) is an open-source operational data system that manages and serves machine learning features to both training and inference pipelines with low latency. It provides a consistent feature definition API, offline serving for training, and online serving via a low-latency store like Redis or DynamoDB, making it a standard choice for feature store implementations.

Exam trap

CompTIA often tests the distinction between ML orchestration tools (Kubeflow, MLflow) and dedicated feature stores (Feast, Tecton), trapping candidates who confuse lifecycle management with feature serving infrastructure.

Full explanation →
417
MCQhard

A company deploys a deep learning model for real-time object detection in autonomous vehicles. The model was trained on high-end GPUs but needs to run on edge devices with limited computational resources. Which technique is most effective for reducing model size and inference latency while maintaining acceptable accuracy?

A.Hyperparameter tuning
B.Batch normalization
C.Dropout
D.Quantization
AnswerD

Quantization reduces numerical precision, shrinking model size and improving inference speed.

Why this answer

Quantization reduces the precision of the model's weights and activations (e.g., from 32-bit floating point to 8-bit integers), which significantly decreases model size and speeds up inference on edge devices with limited computational resources. This technique directly addresses the constraints of edge deployment while often maintaining acceptable accuracy through careful calibration.

Exam trap

CompTIA AI exams often test the misconception that regularization techniques like dropout or batch normalization can reduce model size or inference latency, when in fact they are training-phase optimizations that do not directly address edge deployment constraints.

How to eliminate wrong answers

Option A is wrong because hyperparameter tuning optimizes training settings (e.g., learning rate, batch size) to improve model convergence, but it does not directly reduce model size or inference latency on edge devices. Option B is wrong because batch normalization normalizes layer inputs during training to stabilize and accelerate training, but it adds computational overhead during inference and does not reduce model size or latency. Option C is wrong because dropout is a regularization technique that randomly drops neurons during training to prevent overfitting, but it is typically disabled during inference and does not reduce model size or inference latency.

Full explanation →
418
Multi-Selecthard

A company is deploying a generative AI application that produces structured JSON output for downstream processing. They want to ensure the output is consistently valid JSON and matches a specific schema. Which THREE techniques should they use? (Select THREE)

Select 3 answers
A.Fine-tune the model on a dataset of JSON outputs
B.Increase the temperature parameter to 1.5
C.Provide few-shot examples of the desired output
D.Include a system prompt specifying the expected JSON schema
E.Use JSON mode (structured output) in the API call
AnswersC, D, E

Correct: few-shot examples help the model understand the exact schema.

Why this answer

Structured output (JSON mode) forces valid JSON, system prompts instruct the model, and few-shot examples demonstrate the required schema.

Full explanation →
419
MCQmedium

A financial institution uses a regression model to predict credit risk. The model has a high R-squared on training data but low R-squared on test data. Which of the following is the most likely cause?

A.The features were not standardized before training.
B.The model is overfitting the training data.
C.The model is underfitting the training data.
D.There is multicollinearity among the input features.
AnswerB

Overfitting explains high training and low test performance.

Why this answer

A high R-squared on training data combined with a low R-squared on test data is the classic symptom of overfitting. The model has memorized noise and specific patterns in the training set rather than learning generalizable relationships, causing poor performance on unseen data.

Exam trap

CompTIA often tests the distinction between overfitting and underfitting by presenting a high training metric with a low test metric, tempting candidates to think the model is 'too good' or that data preprocessing (like standardization) is the fix.

How to eliminate wrong answers

Option A is wrong because feature standardization (scaling) affects convergence speed for some algorithms but does not inherently cause overfitting or the described train-test R-squared gap. Option C is wrong because underfitting would produce low R-squared on both training and test data, not high on training and low on test. Option D is wrong because multicollinearity inflates coefficient variances and can reduce interpretability, but it does not typically cause a large discrepancy between training and test R-squared; it affects both sets similarly.

Full explanation →
420
MCQmedium

Refer to the exhibit. A data scientist is training a neural network and observes the training log above. What is the most likely cause?

A.The model is overfitting
B.The model is underfitting
C.The batch size is too large
D.The learning rate is too high
AnswerD

High learning rate causes the optimizer to overshoot minima, leading to divergence.

Why this answer

The training log shows a loss that initially decreases but then suddenly spikes and oscillates wildly, which is a classic sign of divergence caused by a learning rate that is too high. A high learning rate causes the optimizer to overshoot the minima in the loss landscape, leading to instability and failure to converge.

Exam trap

CompTIA often tests the distinction between overfitting and training instability by showing a loss curve that initially drops then spikes, which candidates may misinterpret as overfitting because they focus on the later oscillations rather than the sudden divergence.

How to eliminate wrong answers

Option A is wrong because overfitting would show training loss decreasing while validation loss increases, not a sudden spike and oscillation in the training loss itself. Option B is wrong because underfitting would show a consistently high loss that fails to decrease significantly, not a pattern of initial decrease followed by divergence. Option C is wrong because a batch size that is too large typically leads to slower convergence or a plateau in loss, not the sudden spikes and oscillations seen in the log.

Full explanation →
421
MCQmedium

Refer to the exhibit. A data scientist defines a model configuration in JSON. Which component is missing from the configuration for a complete machine learning pipeline?

A.Training hyperparameters
B.Data preprocessing steps
C.Model type
D.Evaluation metrics
AnswerB

Preprocessing (scaling, encoding) is missing.

Why this answer

A complete machine learning pipeline must include data preprocessing steps to transform raw data into a format suitable for model training. The JSON configuration defines the model type, evaluation metrics, and training hyperparameters, but omits any specification for data cleaning, normalization, feature encoding, or splitting, which are essential for reproducibility and model performance.

Exam trap

CompTIA often tests the misconception that a model configuration is complete if it includes the model type, hyperparameters, and evaluation metrics, but candidates overlook that data preprocessing is a mandatory pipeline stage for transforming raw data before training.

How to eliminate wrong answers

Option A is wrong because training hyperparameters (e.g., learning rate, batch size) are present in the configuration as part of the model training specification, so they are not missing. Option C is wrong because the model type (e.g., 'neural_network', 'random_forest') is explicitly defined in the JSON under the 'model' key, so it is not missing. Option D is wrong because evaluation metrics (e.g., 'accuracy', 'f1_score') are listed in the configuration under the 'evaluation' section, so they are not missing.

Full explanation →
422
MCQeasy

A company is developing an AI policy. Which of the following should be included to ensure accountability for AI-driven decisions?

A.A set of acceptable use cases for the AI
B.A description of the model architecture used
C.A list of approved training data sources
D.Designated roles for human oversight and decision authority
AnswerD

Clear assignment of human oversight roles ensures that someone is accountable for the AI's outputs.

Why this answer

Option D is correct because accountability for AI-driven decisions requires clear assignment of human roles with oversight and decision authority. This ensures that there is a responsible party who can review, override, or be held liable for the AI's outputs, which is a core principle of AI governance frameworks such as the NIST AI Risk Management Framework.

Exam trap

Candidates often confuse the components of an AI governance policy, such as acceptable use or data sources, with the accountability mechanism of human oversight. The key is to remember that accountability requires designated roles with authority to review and override AI decisions.

How to eliminate wrong answers

Option A is wrong because acceptable use cases define scope, not accountability; they do not assign responsibility for decisions. Option B is wrong because describing the model architecture is a technical documentation detail, not a governance mechanism for accountability. Option C is wrong because listing approved training data sources addresses data provenance and bias, but does not establish who is responsible for the AI's decisions.

Full explanation →
423
Multi-Selecteasy

A data scientist is training a supervised learning model for customer churn prediction. Which TWO types of bias are most likely to affect the model's fairness and accuracy if not addressed?

Select 2 answers
A.Algorithmic bias
B.Selection bias
C.Measurement bias
D.Sampling bias
E.Confirmation bias
AnswersB, C

Selection bias arises when the sample is not representative of the population, leading to skewed predictions.

Why this answer

Selection bias (B) occurs when the training data does not represent the true customer population, e.g., using only data from a specific time period or region, leading to a model that fails to generalize. Measurement bias (C) arises from systematic errors in how features are recorded, such as inconsistent data collection methods across customer segments, which can skew predictions and harm fairness.

Exam trap

CompTIA often tests the distinction between data-level biases (selection, measurement) and human cognitive biases (confirmation bias), so candidates mistakenly pick confirmation bias because it sounds plausible in a data science context.

Full explanation →
424
Multi-Selectmedium

Which THREE are common activation functions used in neural networks? (Choose THREE.)

Select 3 answers
A.ReLU
B.Softmax
C.Sigmoid
D.Linear
E.Tanh
AnswersA, C, E

Rectified Linear Unit is widely used in hidden layers.

Why this answer

ReLU (Rectified Linear Unit) is a common activation function in neural networks because it introduces non-linearity while being computationally efficient. It outputs the input directly if positive, otherwise zero, which helps mitigate the vanishing gradient problem compared to sigmoid or tanh. This makes it a default choice for hidden layers in many deep learning architectures.

Exam trap

CompTIA often tests the distinction between activation functions used in hidden layers versus output layers, so candidates mistakenly select Softmax as a general activation function when it is only appropriate for the final layer in classification tasks.

Full explanation →
425
MCQhard

Based on the exhibit, which action is permitted by this policy?

A.Deploy a new model to an endpoint.
B.Update an existing endpoint.
C.Delete an endpoint.
D.Invoke an endpoint for inference.
AnswerA

The policy explicitly allows creating a new endpoint, which is required to deploy a new model to an endpoint. This action is permitted.

Why this answer

The policy grants permissions to create a new model, create an endpoint configuration, and create an endpoint. It does not allow updating, deleting, or invoking an endpoint. Therefore, deploying a new model (which requires creating a new endpoint) is permitted.

Exam trap

CompTIA often tests the distinction between creating a new resource versus modifying or deleting an existing one, leading candidates to assume that broad permissions like creating an endpoint also cover updates or invocations, which is incorrect in policy evaluation.

How to eliminate wrong answers

Option B is wrong because updating an existing endpoint requires the `sagemaker:UpdateEndpoint` action, which is not listed in the policy. Option C is wrong because deleting an endpoint requires the `sagemaker:DeleteEndpoint` action, which is not granted. Option D is wrong because invoking an endpoint for inference requires the `sagemaker:InvokeEndpoint` action, which is absent from the policy.

Full explanation →
426
MCQmedium

A company is deploying a fraud detection model that must return predictions within 100ms to avoid transaction delays. The team is deciding between batch and real-time inference. Which factor most strongly supports a real-time inference architecture?

A.The model requires large amounts of historical data for each prediction
B.The application requires immediate feedback for each transaction
C.The infrastructure budget is limited and must be optimized
D.The model can be retrained weekly using gathered data
AnswerB

Real-time inference delivers low-latency predictions for each request.

Why this answer

Real-time inference is required when the application must return predictions within strict latency bounds (e.g., 100ms) to avoid transaction delays. The need for immediate feedback per transaction directly aligns with a real-time architecture, where each request is processed individually as it arrives, rather than waiting for a batch window. Batch inference would introduce unacceptable latency because it processes groups of records on a schedule, not on-demand.

Exam trap

CompTIA often tests the misconception that batch inference is always cheaper or more efficient, but the trap here is that latency requirements (under 100ms) force a real-time architecture regardless of cost or data volume.

How to eliminate wrong answers

Option A is wrong because requiring large amounts of historical data for each prediction does not dictate real-time vs. batch; it affects feature engineering and storage, not inference latency. Option C is wrong because limited infrastructure budget typically favors batch inference, which can use cheaper, less scalable resources and process data in bulk, not real-time. Option D is wrong because weekly retraining is a model update frequency concern, unrelated to the inference serving architecture; both batch and real-time systems can support periodic retraining.

Full explanation →
427
Multi-Selecthard

A financial services company needs to deploy an ML model for loan approval that must be explainable to regulators. The model is a gradient boosting ensemble. They need to track experiments, log model parameters, and serve the model with explanations. Which THREE tools from the MLOps ecosystem should they use?

Select 3 answers
A.Apache Kafka
B.Docker Compose
C.Weights & Biases
D.Kubeflow
E.MLflow
AnswersC, D, E

W&B provides experiment logging, hyperparameter tracking, and model visualization.

Why this answer

Weights & Biases (W&B) is correct because it provides experiment tracking, hyperparameter logging, and model versioning, which are essential for the regulatory requirement of explainability and auditability. It integrates directly with gradient boosting frameworks like XGBoost and LightGBM to log parameters and metrics, enabling reproducible ML pipelines.

Exam trap

CompTIA AI+ often tests the distinction between general infrastructure tools (like Kafka or Docker Compose) and purpose-built MLOps tools (like W&B, Kubeflow, and MLflow) that directly address experiment tracking, model serving, and explainability.

Full explanation →
428
MCQmedium

A company uses a third-party AI model for sentiment analysis. They want to create a software bill of materials (SBOM) for this AI system. What is the PRIMARY purpose of an SBOM in this context?

A.To record the model's accuracy on benchmark datasets
B.To list all software components and dependencies used in the AI system
C.To document the model's training hyperparameters
D.To provide a user manual for the AI model
AnswerB

SBOM inventories all software components, aiding in vulnerability management.

Why this answer

The primary purpose of an SBOM for an AI system is to provide a complete inventory of all software components, libraries, and dependencies that make up the system. This is critical for vulnerability management, license compliance, and supply chain risk assessment, especially when third-party AI models are integrated. It does not track performance metrics, training details, or user instructions.

Exam trap

CompTIA often tests the distinction between an SBOM (software inventory for security and compliance) and model documentation (like model cards or datasheets) that cover performance, training, or usage details.

How to eliminate wrong answers

Option A is wrong because recording model accuracy on benchmark datasets is a performance evaluation task, not a component inventory function of an SBOM. Option C is wrong because documenting training hyperparameters pertains to model development and reproducibility, not the software supply chain transparency that an SBOM provides. Option D is wrong because a user manual describes how to operate the model, whereas an SBOM is a machine-readable list of software artifacts and their provenance.

Full explanation →
429
MCQmedium

A data scientist is using a Hugging Face transformer model for a sentiment analysis task. They want to optimize inference latency for a mobile app. Which model format and framework combination is BEST suited for on-device deployment?

A.Convert to TensorFlow Lite (TFLite) and run on the device
B.Use the full PyTorch model with JIT scripting
C.Deploy the model on a cloud endpoint and call via REST API
D.Export to ONNX and use ONNX Runtime with GPU
AnswerA

TFLite is optimized for mobile devices, providing low latency and small binary size.

Why this answer

TensorFlow Lite (TFLite) is specifically designed for on-device machine learning inference on mobile and edge devices. It provides a lightweight runtime, hardware acceleration via delegates (e.g., GPU, NNAPI), and reduced model size through quantization, making it the best choice for optimizing inference latency in a mobile app. Converting a Hugging Face transformer model to TFLite allows the model to run locally without network latency, which is critical for real-time sentiment analysis on a smartphone.

Exam trap

Candidates may mistakenly think that other export formats such as ONNX or PyTorch JIT are equally suitable for mobile deployment, but the correct answer is TFLite because it is specifically designed for on-device inference with quantization and hardware acceleration, while ONNX and PyTorch JIT are primarily optimized for server-side or desktop inference.

How to eliminate wrong answers

Option B is wrong because using a full PyTorch model with JIT scripting does not produce a mobile-optimized runtime; PyTorch Mobile exists but JIT scripting alone lacks the quantization and delegate support that TFLite offers for low-latency on-device inference. Option C is wrong because deploying the model on a cloud endpoint and calling via REST API introduces network latency and dependency on connectivity, which defeats the purpose of on-device deployment for a mobile app. Option D is wrong because exporting to ONNX and using ONNX Runtime with GPU is typically designed for server or desktop environments with dedicated GPUs, not for mobile devices where GPU support is limited and ONNX Runtime Mobile is less mature than TFLite for transformer models.

Full explanation →
430
MCQmedium

An organization implements the above access control policy for its AI model registry. During an audit, the auditor discovers that a data scientist deployed a model to production without authorization. Which of the following is the most likely cause?

A.The ML engineer role has approval required, but the manager approved the deployment
B.The time window condition blocked the deployment, but it was overridden by an administrator
C.The policy allows data scientists to deploy to staging, but they exploited a gap in enforcement to promote to production without the required approval
D.The auditor lacks 'deploy_to_production' permission, so they missed the deployment
AnswerC

The policy lacks technical enforcement of the role-based separation.

Why this answer

Option C is correct because the scenario describes a data scientist who bypassed the required approval step by exploiting a gap in the policy enforcement mechanism. The policy explicitly requires approval for production deployments, but the data scientist was able to promote a model from staging to production without that approval, indicating a flaw in the access control implementation or the separation of duties between staging and production environments.

Exam trap

The trap here is that candidates may assume that a policy allowing deployment to staging inherently prevents promotion to production, but CompTIA tests the understanding that enforcement gaps—such as missing approval checks or role escalation paths—can allow unauthorized actions even when the policy appears restrictive.

How to eliminate wrong answers

Option A is wrong because if the ML engineer role has approval required and the manager approved the deployment, this would be an authorized action, not a violation. Option B is wrong because if the time window condition blocked the deployment and an administrator overrode it, that would be a legitimate override action, not an unauthorized deployment by a data scientist. Option D is wrong because the auditor's lack of 'deploy_to_production' permission is irrelevant to the cause of the unauthorized deployment; the auditor's role is to detect violations, not to perform deployments.

Full explanation →
431
Multi-Selectmedium

Which THREE are common causes of data leakage in machine learning pipelines?

Select 3 answers
A.Using time-based splitting for sequential data
B.Using future information to predict the present
C.Using cross-validation on the entire dataset
D.Applying normalization before splitting data into train and test sets
E.Including features that are directly derived from the target variable
AnswersB, D, E

Using data that would not be available at prediction time is a direct form of leakage.

Why this answer

Option B is correct because using future information to predict the present is a classic form of data leakage. In time series or sequential data, if a model is trained on features that include values from a later time point, it gains access to information that would not be available at prediction time, leading to overly optimistic performance metrics and poor generalization.

Exam trap

CompTIA often tests the distinction between valid data splitting practices and actual leakage causes, so candidates may incorrectly select time-based splitting (Option A) as a leakage cause when it is actually a proper technique for sequential data.

Full explanation →
432
MCQeasy

Refer to the exhibit. The monitoring dashboard for a deployed churn prediction model shows a drift detected flag. However, the error rate and latency are within acceptable ranges. What is the most appropriate immediate action?

A.Trigger automatic retraining using the latest data
B.Roll back to the previous model version immediately
C.Ignore the drift since performance metrics are stable
D.Investigate the type and severity of drift before deciding
AnswerD

Understanding drift (covariate vs concept) informs next steps.

Why this answer

Option D is correct because when drift is detected but performance metrics like error rate and latency are still acceptable, it is important to investigate the type and severity of drift before taking any action. Drift may be benign or may indicate a shift that will eventually degrade performance. Option A is wrong because automatic retraining could be risky if the drift is temporary or benign.

Option B is wrong because rolling back immediately discards potential improvements and could be unnecessary. Option C is wrong because ignoring drift may lead to future degradation.

Full explanation →
433
MCQmedium

A developer is implementing a RAG system and needs to choose a similarity metric for retrieving document chunks. The embedding model produces normalized vectors. Which metric is computationally efficient and equivalent to cosine similarity for normalized vectors?

A.Euclidean distance
B.Hamming distance
C.Manhattan distance
D.Dot product
AnswerD

When vectors are unit normalized, dot product equals cosine similarity, and it is computationally efficient.

Why this answer

For normalized vectors, dot product is equivalent to cosine similarity and is faster to compute.

Full explanation →
434
MCQmedium

A healthcare AI startup is developing a model to predict patient readmission risk. The model will be used to allocate post-discharge resources. Which regulatory framework primarily governs the use of patient data in this scenario?

A.GDPR
B.CCPA
C.HIPAA
D.EU AI Act
AnswerC

HIPAA governs the use and disclosure of protected health information (PHI) by covered entities and business associates in the US.

Why this answer

HIPAA (Health Insurance Portability and Accountability Act) is the correct regulatory framework because the scenario involves a healthcare AI startup using protected health information (PHI) to predict patient readmission risk. HIPAA governs the use, disclosure, and safeguarding of PHI by covered entities and their business associates, which includes AI models processing patient data for post-discharge resource allocation.

Exam trap

CompTIA AI+ often tests the distinction between data privacy regulations (HIPAA, GDPR, CCPA) and AI-specific regulations (EU AI Act). Candidates may incorrectly assume the EU AI Act governs all AI data use, but the underlying data type (healthcare PHI) dictates the primary framework.

How to eliminate wrong answers

Option A is wrong because GDPR is a European Union regulation that applies to personal data of EU residents, but the scenario does not specify that the patients are in the EU or that the startup operates under EU jurisdiction; HIPAA is the primary U.S. healthcare data privacy law. Option B is wrong because CCPA is a California state law focused on consumer privacy and data rights for California residents, not specifically tailored to healthcare data or patient readmission models; it does not preempt HIPAA for protected health information. Option D is wrong because the EU AI Act governs the development and deployment of AI systems based on risk categories, but it does not directly regulate the use of patient data; data privacy for healthcare remains under GDPR or local health data laws, not the AI Act itself.

Full explanation →
435
Multi-Selectmedium

Under the EU AI Act, an AI system used for credit scoring is classified as high-risk. Which THREE obligations apply to the deployer of such a system?

Select 3 answers
A.Register the system with a central EU database
B.Publish the model's source code publicly
C.Provide transparency information to affected individuals
D.Conduct a fundamental rights impact assessment
E.Ensure human oversight of the system's decisions
AnswersC, D, E

Deployers must inform individuals that an AI system is making decisions affecting them.

Why this answer

Option C is correct because Article 13 of the EU AI Act requires deployers of high-risk AI systems to provide clear and meaningful transparency information to affected individuals, including the system's capabilities, limitations, and the logic behind decisions. This obligation ensures that individuals subject to automated credit scoring understand how their data is being used and can exercise their rights under the regulation.

Exam trap

The trap here is that candidates often confuse deployer obligations with provider obligations, mistakenly assigning registration and code publication duties to the deployer instead of the provider.

Full explanation →
436
MCQmedium

A data engineering team is building a pipeline to ingest streaming user activity data, process it in real-time, and store features in a feature store for ML models. Which streaming technology is BEST suited for this real-time data ingestion and processing?

A.Apache Kafka
B.Apache Spark SQL
C.Apache Airflow
D.Apache Hadoop MapReduce
AnswerA

Kafka provides high-throughput, fault-tolerant streaming for real-time data pipelines.

Why this answer

Apache Kafka is the best choice because it is a distributed streaming platform designed for high-throughput, fault-tolerant, real-time data ingestion and processing. It provides publish-subscribe messaging, durable log storage, and stream processing capabilities, making it ideal for ingesting streaming user activity data and feeding it into a feature store for ML models.

Exam trap

CompTIA often tests the distinction between batch and stream processing technologies, and the trap here is that candidates confuse Apache Spark SQL (a batch-oriented SQL engine) with Spark Streaming, or mistake Airflow's scheduling capabilities for real-time ingestion.

How to eliminate wrong answers

Option B (Apache Spark SQL) is wrong because Spark SQL is a module for structured data processing using SQL queries, not a streaming ingestion technology; while Spark Streaming exists, Spark SQL itself is not designed for real-time data ingestion. Option C (Apache Airflow) is wrong because Airflow is a workflow orchestration tool for batch scheduling and DAG management, not a real-time streaming ingestion or processing system. Option D (Apache Hadoop MapReduce) is wrong because MapReduce is a batch processing framework that processes data in large, static batches with high latency, making it unsuitable for real-time streaming ingestion.

Full explanation →
437
Multi-Selecteasy

Which TWO are evaluation metrics for classification problems? (Choose two.)

Select 2 answers
A.Precision
B.Mean Absolute Error
C.R-squared
D.Mean Squared Error
E.Recall
AnswersA, E

Correct: Precision is a classification metric.

Why this answer

Precision is a classification metric that measures the proportion of true positive predictions among all positive predictions made by the model. It is calculated as TP / (TP + FP) and is critical when the cost of false positives is high, such as in spam detection or fraud alert systems.

Exam trap

CompTIA often tests the distinction between classification and regression metrics, and the trap here is that candidates may mistakenly select Mean Absolute Error or Mean Squared Error because they are common evaluation metrics, but they are exclusively used for regression problems, not classification.

Full explanation →
438
MCQhard

A company serves a large language model (LLM) on a Kubernetes cluster. The inference latency is acceptable but the cost is high due to GPU usage. The model is 7 billion parameters and requires 16GB GPU memory. The team wants to reduce cost without increasing latency. Which strategy should they implement?

A.Increase the batch size for inference
B.Add more GPU nodes to distribute the load
C.Switch to CPU-based inference
D.Use model quantization to reduce precision
AnswerD

Quantization reduces model size and memory, enabling more efficient GPU usage.

Why this answer

Model quantization reduces the precision of the model's weights (e.g., from FP32 to INT8), which decreases the GPU memory footprint from 16GB to approximately 4GB for a 7B parameter model. This directly lowers GPU cost per inference while maintaining acceptable latency, as the model can run on fewer or cheaper GPUs without increasing inference time.

Exam trap

CompTIA often tests the misconception that adding more hardware (Option B) or increasing batch size (Option A) always reduces cost, when in fact they increase resource usage and cost; the trap is that candidates overlook memory optimization techniques like quantization as a direct cost-reduction strategy.

How to eliminate wrong answers

Option A is wrong because increasing batch size for inference would increase GPU memory usage and could increase latency due to larger memory transfers, not reduce cost without affecting latency. Option B is wrong because adding more GPU nodes would increase cost, not reduce it, and does not address the high GPU memory usage per inference. Option C is wrong because switching to CPU-based inference would drastically increase latency (often 10-100x slower) due to the lack of parallel processing for large matrix operations, violating the requirement to not increase latency.

Full explanation →
439
MCQeasy

A machine learning engineer needs to choose an algorithm for grouping customers into segments based on purchasing behavior without any labels. Which algorithm should the engineer use?

A.K-means clustering
B.Random forest classifier
C.Linear regression
D.Support vector machine
AnswerA

K-means is unsupervised and groups data based on feature similarity.

Why this answer

K-means clustering is an unsupervised learning algorithm that groups unlabeled data into clusters based on feature similarity, making it ideal for segmenting customers by purchasing behavior without predefined labels. It partitions data into K clusters by minimizing within-cluster variance, which directly addresses the requirement of discovering natural groupings in the data.

Exam trap

Cisco often tests the distinction between supervised and unsupervised learning, and the trap here is that candidates may confuse clustering with classification, picking a supervised algorithm like Random Forest or SVM because they think of 'grouping' as a classification task.

How to eliminate wrong answers

Option B (Random forest classifier) is wrong because it is a supervised ensemble method that requires labeled training data to classify instances, not suitable for unlabeled customer segmentation. Option C (Linear regression) is wrong because it is a supervised regression algorithm used to predict continuous values from labeled data, not for grouping unlabeled data. Option D (Support vector machine) is wrong because it is a supervised classification algorithm that requires labeled data to find a separating hyperplane, and cannot perform unsupervised clustering without modifications.

Full explanation →
440
Multi-Selectmedium

A data scientist needs to store large volumes of unstructured log data for future AI model training. They also need to run SQL-based analytics on the data. Which THREE services are appropriate for this requirement? (Choose 3)

Select 3 answers
A.Pinecone
B.Snowflake
C.BigQuery
D.Amazon S3
E.pgvector
AnswersB, C, D

Snowflake is a data warehouse that supports SQL analytics on structured/semi-structured data.

Why this answer

Snowflake is correct because it is a cloud-native data warehouse that supports both structured and semi-structured data (like JSON, Avro, Parquet) via its VARIANT data type, enabling SQL-based analytics on unstructured log data. It also integrates with cloud storage (e.g., Amazon S3) for storing large volumes of raw logs, making it suitable for AI model training pipelines.

Exam trap

CompTIA AI often tests the distinction between purpose-built databases (vector databases like Pinecone and pgvector) and general-purpose analytics platforms (Snowflake, BigQuery, S3), leading candidates to mistakenly select vector databases for log storage and SQL analytics.

Full explanation →
441
MCQhard

A company uses an LLM API to generate customer support responses. They want to prevent the LLM from generating harmful content, even when users attempt jailbreaking. Which defense is MOST effective at the application layer?

A.Output filtering and content moderation
B.Input validation and sanitization
C.Robust training techniques
D.Rate limiting
AnswerA

Output filtering checks the generated text and blocks harmful content, providing a final safety layer.

Why this answer

Output filtering and content moderation is the most effective defense at the application layer because it directly inspects the LLM's generated response before it reaches the user. This approach can catch and block harmful content that results from successful jailbreaking attempts, which input validation alone cannot prevent since the model may still produce undesirable outputs even with sanitized inputs.

Exam trap

Cisco often tests the misconception that input validation is sufficient for LLM security, but the trap here is that jailbreaking exploits the model's generative capabilities, which can only be reliably mitigated by inspecting the output after generation, not just the input.

How to eliminate wrong answers

Option B is wrong because input validation and sanitization, while useful for preventing injection attacks, cannot stop the LLM from generating harmful content if a jailbreak prompt bypasses these checks; the model's internal behavior is not fully controlled by input filtering. Option C is wrong because robust training techniques (e.g., RLHF or adversarial training) are applied during model development, not at the application layer, and they cannot dynamically adapt to novel jailbreak patterns in real-time. Option D is wrong because rate limiting only controls the frequency of API requests, not the content of the responses; it does nothing to prevent a single successful jailbreak from generating harmful output.

Full explanation →
442
MCQmedium

A company uses a pre-trained language model for a legal document classification task. They have limited labeled data (500 documents). Which strategy is MOST effective for adapting the model to this domain?

A.Use a rule-based keyword matching system instead.
B.Train a new model from scratch on the 500 documents.
C.Apply extensive data augmentation to increase dataset size.
D.Fine-tune the pre-trained model on the 500 labeled documents.
AnswerD

Correct; transfer learning works well with small labeled datasets.

Why this answer

Fine-tuning a pre-trained language model on 500 labeled legal documents is the most effective strategy because it leverages the model's existing knowledge of language structure and general semantics, requiring only a small amount of domain-specific data to adapt to the legal classification task. This approach avoids the high data requirements of training from scratch and outperforms rule-based or augmentation-only methods by directly optimizing the model's weights for the target domain.

Exam trap

CompTIA often tests the misconception that more data is always better (trap of Option C) or that starting from scratch is safer (trap of Option B), when in fact transfer learning via fine-tuning is the standard approach for low-resource NLP tasks.

How to eliminate wrong answers

Option A is wrong because rule-based keyword matching lacks the semantic understanding needed for legal document classification, where context and nuance are critical, and it cannot generalize beyond predefined patterns. Option B is wrong because training a new model from scratch on only 500 documents is insufficient for deep learning models, leading to severe overfitting and poor generalization due to the lack of pre-trained linguistic knowledge. Option C is wrong because extensive data augmentation on only 500 documents may introduce noise and unrealistic variations, and it does not provide the same benefit as leveraging a pre-trained model's learned representations, which already capture rich language patterns.

Full explanation →
443
MCQhard

A financial institution is developing a fraud detection model using historical transaction data. The dataset contains over 10 million records, but only 0.01% of transactions are fraudulent. The current model uses a neural network trained with standard cross-entropy loss, and the team applies random undersampling of the majority class to create a balanced training set. However, the model still produces a high number of false positives (legitimate transactions flagged as fraud) and misses approximately 30% of actual fraud cases. The business requires that at least 95% of frauds be caught, and the false positive rate must be below 1% to avoid overwhelming fraud analysts. The team has limited resources to collect additional data and cannot change the model architecture significantly. Which approach should the team take to best meet the business requirements?

A.Use cost-sensitive learning by assigning a higher misclassification cost to the fraud class.
B.Apply feature selection to remove noisy predictors and then retrain the current model.
C.Switch to an anomaly detection algorithm such as Isolation Forest or One-Class SVM.
D.Collect more transaction data, especially fraudulent examples, to naturally balance the classes.
AnswerA

This directly penalizes false negatives more, encouraging the model to catch more frauds while maintaining a low false positive rate through tuning.

Why this answer

Cost-sensitive learning adjusts the loss function to penalize false negatives more heavily, directly addressing the need to catch more frauds while controlling false positives. Collecting more data is impractical and may not resolve the imbalance. Anomaly detection models treat fraud as outliers but often have high false positive rates in this context.

Feature selection does not inherently solve the imbalance or performance metric trade-off.

Full explanation →
444
MCQmedium

A hospital wants to train a diagnostic AI model using data from multiple hospitals without sharing raw patient data. Which privacy-preserving technique allows collaborative training while keeping data local?

A.Differential privacy
B.Federated learning
C.Data anonymisation
D.Data pseudonymisation
AnswerB

Federated learning trains models on local data and only shares model updates, keeping raw data on-site.

Why this answer

Federated learning is the correct technique because it enables multiple hospitals to collaboratively train a shared diagnostic AI model without exchanging raw patient data. Instead, each hospital trains a local model on its own data, and only encrypted model updates (e.g., gradients or weights) are sent to a central server for aggregation. This keeps all sensitive patient information local, directly addressing the requirement of data locality while still benefiting from collective learning.

Exam trap

CompTIA emphasizes the distinction between techniques that alter data before sharing (e.g., anonymization, pseudonymization) and techniques that keep data local and share only model parameters (federated learning). The trap is assuming that anonymizing or pseudonymizing data satisfies the 'keep data local' requirement, but these still involve data leaving the hospital.

How to eliminate wrong answers

Option A is wrong because differential privacy adds noise to data or model outputs to protect individual privacy, but it does not keep data local; it can be applied to centralized or federated settings, but alone it does not enable collaborative training without sharing raw data. Option C is wrong because data anonymisation removes or masks personally identifiable information (PII) from a dataset, but the anonymised data is still shared with other parties, which violates the requirement to keep raw patient data local. Option D is wrong because data pseudonymisation replaces identifiers with pseudonyms, but the pseudonymised data is still shared and can potentially be re-identified, failing to meet the strict local data constraint.

Full explanation →
445
MCQmedium

A security team is evaluating the risk of adversarial examples against their image classification model. Which characteristic best describes an adversarial example?

A.A naturally occurring image that the model misclassifies due to poor training data
B.An input modified by small, intentional perturbations designed to cause misclassification
C.An image that has been resized incorrectly and appears distorted to the model
D.A corrupted image with missing pixels that the model cannot process
AnswerB

Adversarial examples are intentionally crafted with small perturbations that fool the model.

Why this answer

Option B is correct because an adversarial example is specifically crafted by adding small, often imperceptible perturbations to a legitimate input. These perturbations are designed to exploit the model's decision boundaries, causing it to output an incorrect classification with high confidence. This is a fundamental concept in AI security, highlighting the vulnerability of deep learning models to input manipulation.

Exam trap

This exam often tests the distinction between natural misclassifications (due to data quality or model limitations) and intentionally crafted adversarial perturbations, so candidates mistakenly choose options describing data corruption or preprocessing errors instead of recognizing the key element of deliberate, small-scale manipulation.

How to eliminate wrong answers

Option A is wrong because a naturally occurring image that the model misclassifies due to poor training data is an example of a natural misclassification or distribution shift, not an adversarial example which requires intentional perturbation. Option C is wrong because an incorrectly resized image causing distortion is a preprocessing error or data corruption issue, not a crafted adversarial perturbation. Option D is wrong because a corrupted image with missing pixels is a data integrity problem, not a deliberately engineered input designed to fool the model.

Full explanation →
446
MCQhard

During testing of a customer service chatbot, the team notices that the model sometimes generates plausible-sounding but factually incorrect answers about company policies. Which evaluation approach is BEST to systematically detect and quantify this issue?

A.Regression testing comparing old and new model outputs
B.Unit tests on the data pipeline
C.Integration tests for API calls
D.Evaluation framework with faithfulness and answer relevancy metrics on a held-out test set
AnswerD

An evaluation framework using metrics like faithfulness (whether the answer is supported by the source) and answer relevancy can detect hallucination and quantify model performance.

Why this answer

A comprehensive LLM output evaluation framework using a ground-truth dataset of question-answer pairs and metrics like faithfulness and answer relevancy can detect hallucination systematically.

Full explanation →
447
MCQmedium

A data scientist trains a linear regression model to predict house prices. The model has high bias and low variance. Which action would most likely reduce bias?

A.Apply L2 regularization
B.Increase the training dataset size
C.Add polynomial features
D.Remove irrelevant features
AnswerC

Adding complexity reduces bias but may increase variance.

Why this answer

High bias indicates the model is underfitting the data, meaning it is too simple to capture the underlying patterns. Adding polynomial features increases model complexity by introducing non-linear terms, which allows the linear regression model to better fit the training data and thus reduce bias.

Exam trap

CompTIA often tests the bias-variance tradeoff by making candidates confuse regularization (which reduces variance) with methods that reduce bias, or by implying that more data always fixes underfitting.

How to eliminate wrong answers

Option A is wrong because L2 regularization (Ridge regression) reduces overfitting by penalizing large coefficients, which increases bias to lower variance, making bias worse. Option B is wrong because increasing the training dataset size typically reduces variance (helps with overfitting) but does not address underfitting (high bias) — it may even make bias more apparent. Option D is wrong because removing irrelevant features simplifies the model further, which increases bias and is counterproductive when the goal is to reduce bias.

Full explanation →
448
MCQeasy

A team is deploying a deep learning model for real-time image classification on edge devices with limited computational resources. Which technique would best help reduce model size and inference time without significant accuracy loss?

A.Data augmentation
B.Model pruning and quantization
C.Transfer learning
D.Ensemble learning
AnswerB

Pruning removes redundant weights and quantization reduces precision, decreasing model size and speeding up inference.

Why this answer

Model pruning and quantization directly reduce the number of parameters and the precision of weights (e.g., from 32-bit floats to 8-bit integers), which shrinks the model size and speeds up inference on edge devices. This technique is specifically designed to minimize computational load while preserving accuracy, making it ideal for resource-constrained environments like real-time image classification on edge hardware.

Exam trap

Candidates often mistakenly believe that transfer learning alone reduces model size, but it only reuses weights—the architecture remains unchanged. For resource-constrained edge devices, pruning and quantization are the direct methods for compression and speed optimization.

How to eliminate wrong answers

Option A is wrong because data augmentation increases the diversity of training data to improve generalization, but it does not reduce model size or inference time; in fact, it may increase training time. Option C is wrong because transfer learning reuses a pre-trained model to accelerate training on a new task, but it does not inherently reduce model size or inference speed—the model remains large unless combined with pruning or quantization. Option D is wrong because ensemble learning combines multiple models to improve accuracy, but it multiplies the computational cost and memory footprint, which is counterproductive for edge devices with limited resources.

Full explanation →
449
MCQmedium

A team is training a language model using a large text corpus. They want to ensure the model does not learn biased associations between gender and professions. Which data engineering technique should they apply?

A.Remove all gender-related words from the text
B.Use a pre-trained model that is already debiased
C.Apply adversarial debiasing during training
D.Balance the representation of professions across genders
AnswerD

Balancing ensures the model sees equal examples of each gender across professions, reducing biased correlations.

Why this answer

Option D is correct because balancing the representation of professions across genders in the training data directly addresses the root cause of biased associations. By ensuring that each profession appears with roughly equal frequency for all gender references, the model learns statistical correlations that are fair rather than skewed by imbalanced data. This is a fundamental data engineering technique for bias mitigation, as it prevents the model from encoding spurious correlations between gender and occupation.

Exam trap

CompTIA AI often tests the distinction between data-level interventions (like balancing) and model-level interventions (like adversarial debiasing), trapping candidates who confuse training-time algorithms with data engineering techniques.

How to eliminate wrong answers

Option A is wrong because removing all gender-related words eliminates necessary context for the model to understand language, and it does not prevent the model from learning biased associations from remaining contextual clues (e.g., pronouns in surrounding sentences). Option B is wrong because using a pre-trained model that is already debiased does not guarantee the model will remain unbiased on the specific downstream task or dataset; debiasing is often incomplete and may not transfer to new data distributions. Option C is wrong because adversarial debiasing is a training-time technique that modifies the model's internal representations, not a data engineering technique; it operates on the model architecture and loss function, not on the dataset itself.

Full explanation →
450
MCQeasy

A data analyst is cleaning a dataset and finds that 20% of the values for the 'age' column are missing. Which imputation method is most robust if the data is not normally distributed?

A.Mean imputation
B.Median imputation
C.Mode imputation
D.Remove rows with missing values
AnswerB

Median is robust to non-normal distributions.

Why this answer

Median imputation is the most robust method for handling missing values in the 'age' column when the data is not normally distributed because the median is unaffected by outliers or skewness. Unlike the mean, which is sensitive to extreme values, the median provides a central tendency measure that better represents the typical value in non-normal distributions, preserving the dataset's integrity for downstream modeling.

Exam trap

CompTIA often tests the misconception that mean imputation is always the default or best choice for numerical data, but the trap here is that candidates overlook the importance of distribution shape and outlier sensitivity, leading them to select mean imputation despite the data not being normally distributed.

How to eliminate wrong answers

Option A is wrong because mean imputation assumes a normal distribution and is highly sensitive to outliers, which can introduce bias and distort the dataset's variance when the data is skewed. Option C is wrong because mode imputation is typically used for categorical data, not continuous variables like age, and it can lead to loss of granularity and inaccurate representation of the distribution. Option D is wrong because removing rows with missing values reduces sample size and can introduce selection bias, especially if the missingness is not completely at random, which is inefficient and may degrade model performance.

Full explanation →
Page 5

Page 6 of 14

Page 7

All pages

1234567891011121314

Practice AI0-001 by domain

Target a specific domain to shore up weak areas.

AI Infrastructure and TechnologiesAI SecurityAI Concepts and FoundationsAI Concepts and TechniquesMachine Learning and Deep LearningAI Models and Data EngineeringImplementing AI SolutionsAI Implementation and OperationsAI Security, Ethics and GovernanceAI Governance and Ethics
See all domains with question counts →