AI0-001 AI Security Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Monitor API usage for anomalous patterns

  Why it's wrong here

  Usage monitoring detects anomalous inference behaviour after deployment, but cannot verify whether malicious weights or triggers were embedded in the pre-trained artefact itself. It is tempting because runtime anomaly detection genuinely suits detecting compromised credentials or abuse of a live API, yet backdoor assurance requires scanning or attestation of the model supply chain before use.

• ✗

  Use federated learning to train the model

  Why it's wrong here

  Federated learning distributes training across clients to keep data local; it neither detects nor removes malicious triggers embedded in third-party weights. It is tempting because it addresses supply chain privacy concerns, and it would be the right choice when the requirement is training on decentralised sensitive data rather than verifying model integrity.

• ✗

  Implement differential privacy during training

  Why it's wrong here

  Differential privacy adds noise to training data to limit memorisation and membership inference, so it neither detects nor removes an intentionally implanted trigger. It is tempting because it genuinely protects individual records in sensitive training sets, but backdoor assurance demands provenance verification or scanning of the supplied weights instead.

• ✓

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

  Why this is correct

  An SBOM enumerates every component and dependency in the model artefact, letting the company detect tampered or unauthorised layers before deployment. Verifying it against the supplier's signed manifest directly addresses the backdoor concern, which runtime monitoring or prompt filtering cannot detect in a pre-trained model.