Deep Reinforcement Learning Based Secrecy Performance Optimization for Aerial IRS-Assisted Communications

Intelligent reflecting surface (IRS) introduces additional degrees of freedom and randomness for changing wireless channel properties, which enables programmable wireless environments. The introduction of IRS can enhance the secrecy performance of wireless communication systems from different perspectives. Combining IRS with unmanned aerial vehicle (UAV) as aerial IRS provides new degrees of freedom for improving physical layer security by flexible IRS deployment. This paper investigates the secrecy performance of an IRS-assisted wireless communication system, where the IRS is mounted on a UAV to secure the communication between a base station and a receiver in the presence of one or more passive eavesdroppers. To maximize the secrecy rate of the system, the UAV speed, 3D trajectory and the IRS normal vector are jointly optimized. The coupling of the optimization variables and non-convexity makes it hard to solve the problem directly. Therefore, a DDQN based method is proposed to achieve the optimum secrecy performance. Numerical results demonstrate that the secrecy rate can be greatly improved by jointly optimizing the IRS normal vector and other parameters, which makes certain eavesdroppers located in the coverage blind area of the aerial IRS and the observation angle of remaining eavesdroppers deviated from that of reflected signal.