Data Poisoning Prevention & Detection

[All Actors]
1. Data Governance and Source Validation
a. Define clear metrics for "fit for purpose" data
b. Develop data source vetting and validation procedures
c. Implement data provenance and lineage tracking
d. Implement digital signature validation to verify the integrity and authenticity of models, datasets and software components before use in AI pipelines.

2. Data Processing and Monitoring
a. Establish a process to periodically audit datasets for integrity.
b. Use anomaly detection and drift monitoring to identify deviations, set thresholds for deviations, and implement alerts to notify data owners of potential issues
c. Implement real-time monitoring of data distributions and statistical patterns to detect anomalous changes in training data.
d. Establish baseline metrics and threshold monitoring for early detection of data manipulation attempts.

3. Technical Measures
a. Perform data integrity checks for data used in AI systems to detect alteration, corruption and tampering of data.
b. Maintain detailed audit logs of who accessed, modified, or submitted data, with timestamps and user authentication to trace potential poisoning attempts.
c. Report unusual patterns or unexpected model behaviors promptly to model provider.
d. Conduct periodic reviews of model outputs against expected business outcomes.

1. Verify that infrastructure providers validate data sources ingested into AI training and processing pipelines to prevent malicious or poisoned data introduction.

2. Verify data quality checks are embedded in infrastructure services to detect and filter corrupted or suspicious data during AI workload execution.

3. Verify automated anomaly detection systems monitor data flows and storage for unusual patterns indicating data poisoning.

4. Verify infrastructure supports adversarial training and other resilience techniques by providing appropriate compute and tooling capabilities.

5. Verify infrastructure enforces strict access controls to prevent unauthorized dataset modifications at storage or processing layers.

6. Verify that data encryption protects AI workloads’ data at rest and in transit against unauthorized access or tampering.

7. Verify monitoring and alerting systems detect tampering or poisoning signs at the infrastructure level.

8. Verify documented incident response processes exist for infrastructure-related data poisoning threats, with clear escalation and remediation paths.

9. Verify that infrastructure personnel receive training on recognizing and responding to data poisoning risks in AI workloads.

10. Verify that infrastructure providers deploy automated tools to continuously monitor data integrity and detect anomalies across the AI data pipeline.