Adversarial Sybil attacks against Deep RL based drone trajectory planning

A deep reinforcement learning (DRL) approach is proposed for simultaneously maximizing distinct user coverage and data throughput in the presence of Sybil attacks. Adversaries are assumed to be capable of PHY layer spoofing and their own reinforcement learning (RL) to cause disruption to the drone network. The distribution of users is modelled as a Gaussian Mixture Model (GMM) with a fixed mean to model a stationary user population and a time varying mean to model movement of users. The drone communication is assumed to be over the millimeter-Wave (mmWave) frequency band and the presence of users are determined based on the beacon signals transmitted by them. Adversaries are assumed to attack the drone network by transmitting spoofed beacon signals. The simulation results indicate that DRL based approaches for UAV path planning is able to cover a significant number of users after a few training epochs. They also indicate that the presence of spoofed transmissions by the adversaries results in 10 to 22 percent reduction in the amount of data handled and the number of distinct users covered.