FSISEC13: How do you secure AI/ML models and protect training data?
Financial institutions implementing generative AI must establish comprehensive security controls throughout the AI lifecycle, from data preparation to model deployment and monitoring. This includes protecting training data integrity, securing model development environments, and implementing robust controls for inference to prevent unauthorized access, model manipulation, and data poisoning attacks.
FSISEC13-BP01 Implement comprehensive model security controls
Securing AI/ML models requires implementing multiple layers of protection to maintain model integrity and prevent unauthorized access. Establish least privilege access to foundation model endpoints and implement private network communication between AI components using VPC endpoints or AWS PrivateLink. Use customer-managed encryption keys for model artifacts and training data, implement model versioning with integrity checking mechanisms, and establish secure model storage with strict access controls and audit logging.
FSISEC13-BP02 Protect training data integrity
The integrity of training data directly impacts the security and compliance of AI models. Implement data purification filters to detect harmful inputs, establish data lineage tracking for regulatory compliance, and apply classification schemes for sensitive financial data. Deploy continuous monitoring to detect data poisoning attempts and implement backup and recovery procedures aligned with your organization's data protection strategy.
FSISEC13-BP03 Secure model deployment and inference
Securing deployment and inference stages is critical for preventing unauthorized access and protecting against AI-specific attacks. Implement version-controlled prompt catalogs with security review processes, establish model access controls using IAM policies, and deploy monitoring for anomalous invocation patterns. Implement response filtering mechanisms like Amazon Bedrock Guardrails and secure API gateways with appropriate authentication, authorization, and comprehensive logging.
Resources
Documents
