AI-Enabled Jammer Deception Using Decoy Packets

In this work, we present a learning algorithm for a wireless communications network to transmit decoy packets to counter an adversarial sensing-reactive jammer. As the jammer is required to search across channels for data transmissions, decoy packets can have the effect of stalling the jammer on a particular channel, preventing it from continuing its search and leaving legitimate packets unimpeded. A reinforcement learning algorithm trains a deep neural network with an exploration-exploitation algorithm and experience replay. The state- and action-space and reward function are presented as components of the reinforcement learning framework. Our algorithm is tested with software simulations, modeling ZigBee communications nodes using time-division multiple access for medium access control. A reactive jammer is modeled in the simulation, with the goal of disrupting any detected ZigBee transmissions. A means to measure and distribute the reward function and system state to enable edge-learning in this context is presented as part of the implementation. The results demonstrate the effectiveness of our algorithm in mitigating the jamming attack, outperforming a random decoy strategy by a factor of two.