Low-Complexity Neural Networks for Denoising Imperfect CSI in Physical Layer Security

Channel adaptation physical layer security (PLS) schemes are degraded when the channel state information (CSI) is imperfect. Imperfect CSI is due to factors such as noisy feedback, outdated CSI, etc. In this paper, we propose a low-complexity noisy CSI denoising scheme based on the autoencoder architecture of deep neural networks referred to as DenoiseSec-Net. To further reduce complexity, we then propose a hybrid version (HybDenoiseSecNet) that combines a legacy denoising scheme and a shallow neural network to achieve a similar performance as DenoiseSecNet. Simulation results, in terms of bit error rate (BER), secrecy capacity, and normalized mean squared error (NMSE), show the performance improvement of our proposed scheme compared to conventional denoising schemes. Finally, we study the significant reduction in computational complexity of the proposed scheme compared to another neural network scheme.