Abstract
There is a growing interest in developing deep learning methods to solve many resource management problems in wireless edge computing systems where model-based designs are infeasible. While deep learning is known to be vulnerable to adversarial example attacks, the security risk of learningbased designs in the context of edge computing is not well understood. In this paper, we propose and study a new adversarial example attack, called stealthy interference attack (SIA), in deep reinforcement learning (DRL)-based edge computation offloading systems. In SIA, the attacker exerts a carefully determined level of interference signal to change the input states of the DRL-based policy, thereby fooling the mobile device in selecting a target and compromised edge server for computation offloading while evading detection. Simulation results demonstrate the effectiveness of SIA, and show that our algorithm outperforms existing adversarial machine learning algorithms in terms of a higher attack success probability and a lower power consumption.
