Data Poisoning.

An attack on a machine-learning system in which adversaries inject, alter, or relabel training data so the resulting model behaves incorrectly or contains hidden backdoors. It belongs to the AI & ML Security category of cybersecurity.

An attack on a machine-learning system in which adversaries inject, alter, or relabel training data so the resulting model behaves incorrectly or contains hidden backdoors.

Data poisoning targets the training stage of the ML lifecycle. Attackers manipulate datasets — public web crawls, crowd-sourced labels, fine-tuning corpora, or feedback logs — to bias the model, degrade accuracy, or implant trigger-activated behaviour. Carlini et al. demonstrated in 2023 that even tiny fractions of poisoned web data can corrupt large pre-training corpora. Variants include availability attacks (degrade overall accuracy), targeted attacks (cause specific misclassifications), and backdoor attacks (activate on a chosen trigger). Defences focus on dataset provenance and signing, deduplication, anomaly detection on training data, robust learning algorithms, and continuous evaluation against benchmark and adversarial test sets.

Defences for Data Poisoning typically combine technical controls and operational practices, as detailed in the full definition above.

Common alternative names include: Training data poisoning, Dataset poisoning.