Cisco Releases Open-Source Tool to Track AI Model Provenance
Summary
Cisco’s Model Provenance Kit fingerprints AI models using metadata, tokenizer similarity and weight-level signals. Its compare and scan modes help identify shared lineage and find the closest match in Cisco’s fingerprint database.
Key points
- Cisco released Model Provenance Kit, a Python toolkit and command-line interface for investigating AI model lineage.
- The tool fingerprints models using metadata, tokenizer similarity and weight-level signals, including embedding geometry and direct weight comparisons.
- Compare mode identifies shared lineage between two models; scan mode searches Cisco’s fingerprint database for the closest lineage match.
- The tool targets risks from unverified model origins and changes, including poisoned or vulnerable models and difficulties tracing incidents.
- Cisco says the open-source toolkit is available on GitHub and its base-model fingerprint dataset is on Hugging Face.
Article Details
- Event Type
- Cisco released Model Provenance Kit, an open-source Python toolkit and command-line interface for analyzing AI model lineage.
- Impact
- The tool compares model fingerprints to identify shared lineage or find the closest match in Cisco's fingerprint database. It aims to help organizations trace model origins and assess risks involving poisoned or vulnerable models, manipulation, training bias, licensing, regulatory compliance, and AI supply-chain integrity.