Workstream 1 RFC: Standardizing Model Vulnerabilities

[andrewelizondo]

Standardizing Model Vulnerabilties

Authors:

• @andrewelizondo
• @seanpmorgan
• @iamfaisalkhan

Summary

Currently there are several vendor specific ways to track model vulnerabilities and attack methods. We should standardize this into an industry standard set of vulnerabilities related to specific models and how to detect them.

Priority

• P1: This is important to include in the next release from this workstream.

Level of Effort

• Medium: This will take a week or two to document.

Drawbacks

• Model vulnerabilities are much more rare than typical software vulnerabilities so they may still be edge use cases.

Alternatives

• Mix between CVE, CWE, Vendor specific methods.

Reference Material & Prior Art

• https://research.jfrog.com/model-threats/h5-lambda/
• https://protectai.com/insights/knowledge-base/deserialization-threats/PAIT-KERAS-101

Unresolved questions

• What mechanisms are missing today that make this inconsistent?