Dataset Provenance & Poisoning Prevention
Objective
Protect training investment ($50k-$500k per model) from backdoored data.
Control / requirement
Hash verification + source allowlist + poisoning detection.
Business impact
Failure to implement this control creates risk of poisoned training data causing model behaviour manipulation, with estimated financial exposure of $500k-$5M based on incident data from 2024-2025 AI supply chain attacks.
Validation approach
Test ID: D1-CTL-01-VTS-001 Test Type: Automated Test Design: Provide dataset metadata with hash and source; execute poisoning detection scan against GAISSF™ Benchmark Dataset (Category: Data Poisoning) (open-source options: Giskard, CleanLab) Execution Steps: 1. Load dataset metadata (hash, source URL, date) 2. Verify hash matches known-good registry 3. Execute poisoning detection scan: pytest tests/d1_model_integrity/test_dataset_provenance.py -v # oda3-gaissf-vts 4. Generate report with poisoning_score Pass Criteria: hash_verified = True; source_in_allowlist = True; poisoning_score = 0; test_coverage >= 1000 samples Independent Verification: Auditor re-runs poisoning detection using GAISSF™ test harness (open-source options: Giskard, CleanLab) with auditor-controlled dataset sample. Result hash compared to organisation's submission.
Expected evidence
Not separately specified in the available source.
Mapping and source
Not separately specified in the available source.
Implementation guidance
Translate the requirement into system-specific procedures, responsible roles, technical configurations and review conditions. Retain evidence showing both design and operating performance. Where the source provides no separate implementation instruction, do not infer that a single technical mechanism is sufficient.
Assessment considerations
- Confirm scope and applicability.
- Inspect control design and responsible ownership.
- Test representative operation and adverse conditions where appropriate.
- Evaluate evidence provenance, completeness and contradictory evidence.
- Record limitations and notably absent outcomes.