AI Security Atlas › All entries › Model Supply Chain & Training Integrity
PII Scrubbing & Privacy-Preserving Training
Training data is scrubbed and deduplicated, sensitive sets use differential privacy, and models are tested for memorization before release.
Open in the interactive atlas →Impact
Reduces what the model can memorize and catches leakage before release.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Unscrubbed Fine-Tuning Data (Medium), Training Data Memorization Leakage (High).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Unscrubbed Fine-Tuning Data Vulnerability · Medium severity
- Training Data Memorization Leakage Threat · High severity
Builds on
- Dataset Provenance & Poison Screening Guardrail · Fixes high risks
Remediation
- Redact PII and secrets with typed placeholders before training.
- Deduplicate at document and near-duplicate level.
- Use differentially private fine-tuning for sensitive datasets.
How to verify
- Canary extraction tests run as a release gate.
- Scrubbing reports are archived with each dataset version.
Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM02:2025 Sensitive Information Disclosure
MITRE ATLAS
NIST AI RMF 1.0
- MEASURE 2.10 Privacy risk of the AI system is examined and documented