Zum Hauptinhalt springen
Glossary

Model poisoning

Updated on 1 min

Model poisoning covers attacks that target the model itself: its weights, pre-trained components or the files in which a model is distributed. Typical cases are a model from a public repository that deliberately misbehaves on certain inputs, or a model file that executes malicious code when loaded. In federated learning, the term also covers manipulated model updates sent by individual participants. The organization’s own training data stay untouched, which sets model poisoning apart from data poisoning.

The AI Act requires high-risk AI systems to include safeguards against manipulated pre-trained components (Art. 15(5)). In practice this means obtaining models only from vetted sources, verifying checksums and signatures, using file formats that cannot carry executable code and treating models like any other software in supplier management. Background in T09 Model poisoning.

No Cookies!

This wiki collects nothing, bakes nothing and leaves nothing behind. There's nothing to consent to. Privacy doesn't get better than this.