Physical-layer secret key generation based on domain-adversarial training of autoencoder for spatial correlated channels.

As a novel approach to enhance information security, physical-layer key generation is based on the channel reciprocity and spatial decorrelation of the wireless channels between two legitimate sides. Due to the half-duplex mode of communication systems, the channel responses detected by the two sides are not exactly reciprocal. Also, the eavesdropper may be extremely close to the legitimate side, and information leakage could arise due to the spatial correlation in this case. To solve this problem, this paper proposes an efficient physical-layer key generation scheme based on the combination of an autoencoder and Domain-Adversarial Training of Neural Networks (DANN), i.e., Domain-Adversarial Training of Autoencoder (DAAE). DAAE extracts the reciprocal channel features of the legitimate sides while maximizing the feature difference on the eavesdropper. Simulation experiments are conducted to verify the effectiveness of DAAE, and the experimental results show that based on DAAE, the correlation of extracted features between the eavesdropper and its adjacent node is weakened without significant influence on the legitimate sides. Meanwhile, after quantization, the primary key shows a good consistency rate and randomness. The proposed method has a good potential for industrial applications. [ABSTRACT FROM AUTHOR]