Task-Oriented Secure Semantic Communication System based on Generative Adversarial Network
Kai Tong, Han Hu, Wei Wu, Fuhui Zhou, Xiaoming He, Yongan Guo · 2025
In the six generation (6 G) era, semantic communication is regarded as a highly promising communication paradigm, which is expected to break through the performance bottlenecks of traditional communications and drive the innovation of communication technologies. As an emerging communication paradigm, semantic communication is vulnerable to eavesdropping threats. Currently, most secure semantic communication works focus on the image reconstruction security at the receiver, ignoring the performance security of downstream artificial intelligence (AI) tasks. In this paper, we propose a secure semantic communication system based on generative adversarial network (GAN). Specifically, we integrate generative adversarial network into the semantic communication architecture to reduce the image reconstruction quality of the eavesdropper. Then, we design a secure semantic feature loss function based on the semantic-level information required for the downstream AI task to mitigate the eavesdropper’s capability to preserve task-related semantic information. In addition, we propose a novel FPSNR metric based on the traditional Peak Signal-to-Noise Ratio (PSNR) metric to better evaluate the system secure performance from semantic perspective. Finally, the simulation results demonstrate that the proposed scheme can significantly ensure the performance security of the downstream AI task.