Learning attack mechanisms in Wireless Sensor Networks using Markov Decision Processes

Abstract

In this work, we identify two related problems that arise in many Wireless Sensor Networks defense mechanisms: the problem of ad-hoc defense and the problem of optimality. These problems open the door to attacks that could severely affect the performance of defense mechanisms. In this work, we use Markov Decision Processes as framework to model an attacker that is able to exploit these two problems. This allows us to model a defense mechanism theoretically and to obtain the performance of an attack against it, as well as to obtain the optimal attack against the defense mechanism - i.e., the attack that harms the most the defense mechanism. We also make use of Deep Reinforcement Learning tools, showing that they can be used by an intelligent attacker to successfully exploit a possibly unknown defense mechanism, providing a compromise between attack results and computational cost. We test our approach by thoroughly studying a Cooperative Spectrum Sensing attack, which we use to illustrate the framework proposed and to highlight their strengths and weaknesses.