Cisco Releases Model Provenance Kit for AI Model Verification

Cisco has released the Model Provenance Kit, an open-source Python toolkit designed to verify the origins and authenticity of AI models, according to a company blog post. The tool addresses AI supply chain security risks by analyzing model metadata and learned parameters to determine whether models share a common training lineage.

The toolkit uses a two-stage comparison process. The first stage rapidly compares architectural metadata and configurations to identify related models. When metadata is ambiguous, a second stage evaluates five weight-level signals: Embedding Anchor Similarity, Embedding Norm Distribution, Norm Layer Fingerprint, Layer Energy Profile, and Weight-Value Cosine. These signals generate a composite provenance score that reflects shared lineage between models.

Model Provenance Kit includes two operational modes. Compare mode allows users to analyze two specific models, while scan mode matches a model against a fingerprint database containing around 150 base models. Testing on 111 model pairs showed 100 percent recall for identifying derivatives and no false positives when models only shared tokenizers.

The toolkit runs on CPU and scales with model size. It is available on GitHub and the Hugging Face platform as part of Cisco Foundation AI's open-source initiative for AI supply chain security.