Which three security measures could be applied in different ML workflow stages to defend them against malicious activities? (Select three.)

A defense-in-depth strategy for ML workflows involves applying security controls at multiple stages. Launching instances in a Virtual Private Cloud (VPC) provides network isolation for training and deployment environments. Data encryption protects sensitive data at rest (in storage) and in transit across all workflow stages. Using a secrets manager to protect credentials (like API keys and database passwords) is crucial for securing access to data and ML artifacts, preventing unauthorized use by hardcoding secrets. These three measures cover network, data, and access control security.

A. Disabling logging is a counter-security measure; logging is essential for auditing, monitoring for threats, and incident response.
C. Monitoring model degradation is primarily for performance and accuracy (MLOps), though it can indirectly detect some attacks. It's less of a direct security defense than the chosen options.
E. Using maximum privilege violates the principle of least privilege, a fundamental security best practice that minimizes potential damage from a compromised account.

1. Amazon Web Services (AWS). (2022). Security Best Practices for Amazon SageMaker. Section: "Network and Data Protection." This official documentation recommends using VPCs for network isolation (p. 8) and encryption for data at rest and in transit (p. 10).

2. Google Cloud. (2023). AI/ML security best practices. Section: "Secure your data and artifacts." This documentation explicitly advises using encryption for data and using Secret Manager to store secrets like API keys.

3. Microsoft Azure. (2023). Enterprise security and governance for Azure Machine Learning. Section: "Network security." This documentation details the use of Virtual Networks (VNets), Azure's equivalent of VPCs, to secure ML workspaces. It also covers identity and access management, aligning with the use of a secrets manager.