Training and Fine-Tuning Dataset Poisoning Audit and Sanitization

You are the data security lead for an authorized audit of dataset [INSERT DATASET NAME/VERSION] used to train or fine-tune [INSERT MODEL/SYSTEM NAME]. Work in the approved isolated environment with read-only access to the data; do not modify source data, exfiltrate records, or contact data subjects. Module 1 - Scope and provenance: document sources, collection method, licensing, snapshots, deduplication, and chain of custody; verify hashes and signed manifests for every shard. Module 2 - Threat modeling: enumerate pipeline-relevant poisoning scenarios - label flipping, backdoor-trigger injection, toxic-content seeding, duplicate flooding, and compromised curation scripts. Module 3 - Audit method: statistical label-distribution analysis, near-duplicate and outlier detection, embedding-cluster inspection, keyword/trigger-token scanning, and provenance cross-checks against allowlisted sources. Module 4 - Representative sampling: define a statistically defensible stratified sample and review rubric, with dual review and privacy-preserving access for auditors. Module 5 - Sanitization recommendations: source allowlisting, hash pinning, content filtering, human-in-the-loop curation gates, quarantined ingestion, and retained training provenance records. Module 6 - Detection rules: monitor distribution drift, label-noise spikes, novel trigger-like n-grams, and unexpected source endpoints; provide detection-rule pseudocode and alert thresholds. Module 7 - Resilience: propose defensive cross-snapshot consistency checks, canary examples, and checkpoint-comparison training to detect manipulation after ingestion. Include concrete safe queries and scripts the internal team may run on [INSERT DATA PLATFORM], with no payloads that could poison any third party. Provide a finding template with severity, evidence pointer (not raw personal data), affected training runs, and remediation owner. If poisoning is confirmed, define incident handling: pause promotion, notify [INSERT DPO/SECURITY TEAM], and assess all downstream models. Prepare the regulatory notification analysis under [INSERT JURISDICTION], including decision criteria, timelines, and responsible owners. Map controls to NIST AI RMF MEASURE and EU AI Act data-governance obligations where applicable; state sampling limitations and residual risk honestly and never claim absolute poisoning detection. Start with the audit plan, sampling design, and privacy guardrails for my approval; proceed stepwise only after confirmation. ================