Security Improvement for Deep Learning-Based Semantic Communication Systems

Semantic communication is becoming increasingly popular for wireless image transmission due to its superior communication efficiency. However, current deep learning-based semantic systems designed for semantic communication, though efficient, remain vulnerable to eavesdropping and often overlook security measures at the physical layer. To address this issue, this paper presents a deep learning-based system with joint source-channel coding (JSCC) and cyclical consistent generative adversarial network that enhances the security of semantic communication systems. We also design a convolutional neural network for the encoder and decoder which is trained to extract and transmit semantic features while minimizing the risk of privacy leakage. The artificial noise at the physical layer is employed at the source to degrade the eavesdropping ability of the eavesdropper. We show through experiments that under the artificial noise strategy, the legitimate user achieves a higher structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) than the eavesdropper. Moreover, the semantic system with JSCC offers better SSIM and PSNR than the separated source and channel coding models while preserving the confidentiality of semantic information during wireless transmission. This enhanced security framework opens new opportunities for secure and reliable communication of semantic information in diverse applications.