Cisco Unveils Free Open-Source Tool for Tracking AI Model Origins

Post Views: 4

Cisco Releases Open Source Tool for Tracking AI Model Provenance

Cisco has launched an open-source tool called Model Provenance Kit, aiming to address concerns around model poisoning, bias, and regulation compliance.

Background

The proliferation of pre-trained AI models has revolutionized the field of natural language processing and computer vision, but also poses significant security risks due to the reliance on third-party models.

The lineage of models becomes increasingly difficult to track as they are fine-tuned, distilled, and repackaged.
This obscurity hinders effective response and remediation efforts when a compromised model is detected.

Model Provenance Kit

Model Provenance Kit generates a unique fingerprint for each model based on metadata signals, tokenizer similarity, and weight-level identity signals.

It operates in two modes:
Compare: Users can identify shared lineage between models.
Scan: It finds the closest lineage for a given model by comparing its fingerprint against a comprehensive database of fingerprints.

According to Cisco, the release of Model Provenance Kit represents a crucial step toward establishing an evidence-based approach to model provenance. By providing a transparent and accountable framework for tracking AI model lineage, organizations can ensure the reliability and trustworthiness of their AI-powered systems.

Availability

The Model Provenance Kit is available on GitHub, along with a comprehensive dataset of base model fingerprints on Hugging Face. This open-source initiative encourages collaboration among developers and researchers, fostering a community-driven effort to enhance the security and accountability of AI models.