Summary

In this chapter, we covered a lot of material on supply chain risks. We looked at traditional third-party vulnerability management and supply chain risks from an AI development perspective. We also looked at the ability of vulnerable packages to help stage stealthily adversarial AI attacks, such as poisoning or tampering. We looked into mitigating this risk with enhanced strategies such as private package repositories and curated package repository lists.

We extended our discussion to the bloodline of AI, models, and data. We demonstrated the additional risks that supply chain attacks bring, especially to poisoned models. We discussed checks and tests that we can apply and, more importantly, the role of provenance, governance, and lineage to reduce our risks. Finally, we looked at simple examples of how to roll these out using private model repositories and MLOps platforms, such as MLflow.

We will look more into these topics and how they all fit together in the MLSecOps...