Data Poisoning

Data poisoning is an attack that changes or inserts examples in a model’s training pipeline to corrupt what the system learns. An attacker may aim to reduce overall accuracy, create a hidden backdoor triggered by a specific pattern, bias decisions against a target, or influence a generative model’s responses. Poisoning can enter through public data collection, compromised suppliers, user feedback, labels, retrieval indexes, or repeated fine-tuning. Defenses include provenance tracking, access control, anomaly detection, robust training, duplicate analysis, trusted validation sets, and review of unexpected behavior. The attack is difficult to diagnose because harmful examples may appear ordinary and their effect may emerge only after deployment.