Deep Learning Aided Physical-Layer Security: The Security Versus Reliability Trade-Off

Abstract

This paper considers a communication system whose source can learn from channel-related data, thereby making a suitable choice of system parameters for security improvement. The security of the communication system is optimized using deep neural networks (DNNs). More explicitly, the associated security vs reliability trade-off problem is characterized in terms of the symbol error probabilities and the discrete-input continuous-output memoryless channel (DCMC) capacities. A pair of loss functions were defined by relying on the Lagrangian and on the monotonic-function based techniques. These were then used for managing the learning/training process of the DNNs for finding near-optimal solutions to the associated non-convex problem. The Lagrangian technique was shown to approach the performance of the exhaustive search. We concluded by characterizing the security vs reliability trade-off in terms of the intercept probability vs the outage probability.