Defenses and mitigations
The implications of successful model extraction attacks include unauthorized replication of proprietary models, financial losses, and competitive disadvantages. Defending against these attacks requires a multifaceted approach to investing in prevention and detection. It also entails additional controls to help identify and recover cloned models. Recovery in this context is not necessarily a physical recovery but, as we will see, a set of actions that will help us counter the impact of an adversarial model extraction.
Prevention measures
As mentioned in Chapter 3, preventing model extraction attacks requires having measures in traditional cybersecurity controls, including firewalls, encryption, access control, and more specialized adversarial AI defenses such as model hardening and adversarial training. We will review these defenses using a defense-in-depth philosophy while tracing an attack’s kill chain. Remember that an extraction attacker works...
