Advanced MLSecOps with SBOMs

We have integrated model evaluation in our MLSecOps pipeline, but it is still nowhere near the thorough vulnerability testing we use for software packages. This is because vulnerability reporting and scanning are still in their infancy in AI. Databases such as airisk.io, now owned by MITRE, and standards such as the OWASP Cyclone DX ML SBOM (Signed Bill of Materials) are initiatives that will transform the MLSecOps space, allowing us to apply similar diligence to AI artifacts, including data and models.

The following diagram summarizes the vision of a robust MLSecOps pipeline to secure models:

Figure 18.10 – A reference MLSecOps pipeline

It uses safety benchmarks, such as Decoding Trust to evaluate against Trustworthy AI metrics such as toxicity, bias, and so on. We use this to create an SBOM using the Cyclone DX format and signed attestations, as we discussed in Chapter 6.

Note

You may not want to run all our...