Hands-On Artificial Intelligence for Cybersecurity

Black-box attacks with RL

In the paper mentioned previously, an RL model is implemented to carry out a black-box attack against a malware detector, with the aim of evading a static Windows Portable Executable (PE) malware classifier.

The attack scenario includes the following elements:

• The RL model consists of an agent and an environment.
• The agent iteratively chooses an action A to execute.
• Each action A is associated with a change in state space S.
• Every change of state is associated with feedback from the environment, in the form of a scalar award.

The feedback and their scalar awards are then fed back to the agent. The agent determines his next action based on such feedback, following a strategy, that is, an objective function that ...