Defining and Measuring Deception in Sequential Decision Systems: Application to Network Defense

Concealing true targets plays a crucial role in preventing adversaries from launching effective attacks. Deception techniques enhance security by misleading and delaying adversaries. This paper presents a framework for defining and measuring deception in sequential decision systems modeled as Markov decision processes with task constraints. We introduce three types of deception: diversionary, targeted, and equivocal, which are implemented by synthesizing deceptive policies using constrained optimization. Adversaries are modeled with inverse reinforcement learning to infer system goals. Numerical results demonstrate the effectiveness of our deception techniques in misleading adversaries and enhancing system security.