Evading ML detectors

In Chapter 8, GANs – Attacks and Defenses, we showed how to use Generative Adversarial Networks (GANs) to deceive detection algorithms. Now, we will see that, it is not only GANs that pose a threat to our AI-based cybersecurity solutions, but more generally, it is possible to exploit Reinforcement Learning (RL) to render our detection tools ineffective.

To understand how, we need to briefly introduce the fundamental concepts of RL.

Understanding RL

Compared to the various forms of AI, RL is characterized by implementing a trial and error fashion of automated learning. In fact, the RL algorithms adapt their learning processes based on the feedback obtained from the environment. This feedback can...