Question

A financial services company trains a fraud-detection ML model on transaction data continuously ingested from multiple partner banks. A security engineer discovers that an attacker with access to one partner's feed has been submitting subtly mislabeled transactions over several months, gradually degrading the model's ability to flag a specific fraud pattern. Which control would most effectively mitigate this type of attack against the training pipeline?

Accepted Answer

Implement data provenance tracking with statistical anomaly detection and validation of incoming training samples before ingestion

Answer

Encrypt the training dataset at rest using an HSM-backed key hierarchy

Answer

Deploy a web application firewall in front of the model's inference API endpoint

Answer

Require multi-factor authentication for all data scientists accessing the model registry