Diffusion Model Based Secure Semantic Communications with Adversarial Purification

A new communication paradigm based on deep learning, known as semantic communication, is driving research into end-to-end data transmission in tasks such as image classification and reconstruction. However, the issue of security stemming from semantic perturbations remains largely unexplored, leading to vulnerabilities in semantic communication systems. In this paper, we propose a secure semantic communication system that utilizes the diffusion model to tackle this issue. The secure semantic communication system proposed in this paper mitigates perturbations caused by semantic-oriented attacks by employing a diffusing process at the sender side and a denoising process at the receiver side. Simulation results indicate that, compared to conventional methods, the proposed secure semantic communication system exhibits superior robustness and accuracy across various channel conditions.