Question

A financial services company fine-tunes a foundation model in Amazon Bedrock using a custom dataset that is periodically supplemented with user-submitted feedback stored in an S3 bucket. During a security review, the governance team is concerned that malicious actors could inject crafted examples into the feedback pipeline to bias or backdoor the model (training data poisoning). Which approach BEST mitigates this risk before the data is used for fine-tuning?

Accepted Answer

Implement a data validation and curation stage that screens submitted examples (e.g., anomaly detection, human review, provenance checks) and restricts write access to the training dataset with least-privilege IAM before creating the customization job.

Answer

Enable Amazon Bedrock Guardrails on the fine-tuning job so poisoned examples are filtered during training.

Answer

Encrypt the S3 training bucket with a customer-managed KMS key so injected data cannot be read by the fine-tuning process.

Answer

Enable CloudTrail data events on the S3 bucket to detect and automatically remove poisoned records after fine-tuning completes.