Secure and Robust Joint Source-Channel Coding With Semantic Clustering and Adversarial Purification

Vision Transformer (ViT) and Swin Transformer have attracted significant attention in the fields of deep learning-based joint source-channel coding (JSCC) and semantic communication due to their excellent visual modeling capabilities. However, we have identified three challenges in these methods: (1) Quadratic complexity of global attention in ViT introduces a substantial computational burden; (2) Token grouping based on windows in Swin Transformer reduces computational complexity, but this approach neglects the semantic information within the tokens. It may divide tokens with similar semantics into different windows or divide tokens with different semantics into the same window, leading to reduced effectiveness; (3) Vulnerability to adversarial attacks in deep learning-based JSCC poses a significant threat to semantic communication, compromising the security and robustness of the system. To address these challenges, we propose a semantic clustering and adversarial purification-based JSCC (SCAPJSCC) scheme. It not only reduces the computational complexity of the self-attention mechanism but also preserves and leverages the semantic information inherent in images. Furthermore, we introduce a plug-and-play adversarial purification module on the receiver, enhancing the robustness and security against adversarial attacks at both the transmitter and the communication channel. Experimental results demonstrate that SCAPJSCC outperforms the state-of-the-art method SwinJSCC, achieving more effective semantic modeling of image information, and stronger resilience to various adversarial attacks.