A model signature is a cryptographic mark attached to a model artifact, such as weights, a checkpoint, or a packaged deployment file. It allows a verifier to check two things: that the artifact has not been changed since it was signed, and that it came from a trusted source. In practice, this is similar to code signing for software, but applied to machine learning assets.
Model signatures matter because AI supply chains are easy to tamper with. Attackers may try to replace a model with a poisoned version, insert a backdoored checkpoint, or swap an approved artifact with a lookalike file. A valid signature helps defenders detect these changes during storage, transfer, and deployment. It is a strong integrity control, but it does not prove the model is safe, accurate, or free of harmful behavior. Security teams use signatures alongside provenance records, access controls, and runtime validation to reduce the risk of model tampering and unauthorized substitution.