Training Data Poisoning
Data poisoning involves injecting malicious data into ML training sets to corrupt model behavior. This can cause the model to make incorrect predictions, include backdoors that activate on specific triggers, or degrade overall performance.
Overview
Training data poisoning is a supply chain attack on machine learning systems. By injecting carefully crafted malicious samples into the training dataset, attackers can cause the model to learn incorrect behaviors, introduce backdoors that activate on specific trigger patterns, or gradually degrade model performance.
Impact
Corrupted model predictions, backdoor activation, degraded model accuracy, biased outputs, supply chain compromise.
Remediation
Validate and sanitize training data, implement data provenance tracking, use robust training techniques, monitor model performance for drift, and maintain data quality pipelines.
Validate the provenance and integrity of all training data
Implement statistical outlier detection on training datasets
Use robust training techniques resistant to poisoning
Monitor model predictions for drift and anomalies
Maintain versioned, auditable training data pipelines
Related SuperClue scanner checks
These scanner checks look for signs of Training Data Poisoning. What a scan can confirm from outside depends on your application.
Scan a domain you have verified with up to 350 modules, paid in credits.
See how website scanning works