← All articles
AI security

Defensive checks for training data poisoning

Use dataset provenance and review checkpoints to spot suspicious changes before training.

Track the origin of each data batch

Record where a batch came from, when it was imported, and the transformations applied before training. This makes it possible to isolate an unexpected batch and reproduce the earlier dataset. Pay particular attention to feeds where new data enters without manual review.

Investigate changes before release

Compare a new version with the previous accepted one. Look for duplicate clusters, unusual label shifts, or a single source contributing most of a rare class. Treat these as review signals and inspect samples in context. Hold batches with unexplained changes for review before using them in training.

Establish an accepted dataset baseline

Keep the origin and import time for each batch along with label schema and preprocessing versions. Record which sources are expected to supply each category. For instance, a new camera feed may legitimately contribute more night-time images. A sudden shift is easier to explain when you know which feed changed and who approved it.

Inspect suspicious changes in context

Compare batch sizes, label frequencies, and duplicate clusters with the previous release. If one source suddenly supplies most examples of a rare category, inspect a sample of those examples and check them against the guidelines. A distribution shift can be legitimate. Use the review to explain it and document the evidence before the batch passes through to training.

Put release gates at ingestion and training

Validate schemas and required metadata at ingestion. Hold unexpected or unverifiable batches for separate review before joining them with approved data. Pin the training job to an approved dataset version so a newly uploaded batch is never included silently. Keep intermediate artifacts and the accepted manifest long enough to reproduce an earlier run if a later data quality issue is discovered.

Investigate model changes against dataset history

If a model suddenly degrades on a stable evaluation slice, compare model settings, code, and training data versions before assuming a cause. Inspect which sources added or replaced examples for the affected task. If you find a suspect batch, run a controlled comparison using the accepted previous version or a reviewed replacement. Record what the comparison establishes and what remains uncertain.