Robust Semantic Communication via Adversarial Training
Kai Wei, Renjie Xie, Wei Xu, Zhaohua Lu, Huahua Xiao · IEEE Transactions on Vehicular Technology · 2025
The increasing demand for transmitting massive data in intelligent communication systems has garnered much attention on joint source-channel coding (JSCC) and semantic communication. Existing neural network (NN)-based semantic communication solutions predominantly focus on specific channel environments, which limits their adaptability to varying channel conditions. To address this challenge, we present a robust semantic communication approach that leverages adversarial (ADV) training to broaden its adaptability across diverse unseen channel conditions. While ADV training has been previously explored in semantic communications, its application in improving the robustness of JSCC in varying channel conditions remains underdeveloped. Specifically, our method enhances the robustness of semantic communication systems under varying channel conditions by incorporating ADV training, subjecting NNs to carefully crafted ADV perturbations that mimic diverse channel distortions. This enables the NN to perform more effectively when encountering changes in channel states. Simulation experiments demonstrate that integrating ADV training yields a substantial performance improvement of over 10% when channel states change, without increasing NN complexity. Notably, unlike existing methods that concentrate on specific channel models, our approach offers a more generalized solution that effectively adapts to a wide range of real-world channel conditions. This study contributes to the development of robust semantic communication systems empowered by ADV training, capable of adapting to diverse channel conditions in real-world scenarios.