Data Poisoning Attacks

Data poisoning studies how malicious or corrupted examples can alter a model during training, from degrading overall accuracy to implanting a backdoor that activates only under a chosen trigger. Attacks include label flipping, clean-label and trigger-based poisoning, malicious client updates in federated learning, and poisoned documents or preferences used by language models and retrieval systems. Defenses use data provenance and filtering, robust aggregation, representation and influence analysis, certified training, and post-training unlearning. A central challenge is that small, carefully targeted poisons can preserve normal benchmark performance, while defenses often depend on the model, data pipeline, and attack type.