A Physical Layer Security Communication Mechanism Based on End-to-End Learning

Abstract

Due to the randomness of wireless channels, it is possible to achieve secure communications by designing communication systems according to the characteristics of wireless channels. In this work, a physical layer security communication mechanism based on end-to-end learning is introduced. The proposed mechanism achieves secure communications by designing the loss function of end-to-end learning. We evaluate secrecy performances of the proposed mechanism through block error rate (BLER) under three classic circumstances which different decoding models are applied by the eavesdropper. Self-training decoding model, the legitimate decoding model and unsupervised learning (K-means clustering) model are alternative for the eavesdropper. The experiments results show that K-means clustering model can get the best BLER performance for the eavesdropper with some prior information, but its BLER is still higher than 0.5. Even if the eavesdropper has a much higher than the legitimate users, there is no improvement in its BLER traditional physical layer security E s /N 0 . Compared to methods, the proposed mechanism can achieve secure communications without extra key extraction.