基于信息瓶颈准则约束的对抗鲁棒语义通信方法

Adversarial robust semantic communication method based on information bottleneck criterion constraint

摘要:

基于深度学习的语义通信旨在传递用户意图和语义信息,有望成为6G网络“内生智能”架构的重要技术支撑,但语义通信系统的对抗鲁棒性及其安全性尚未得到充分研究。为此,提出了一种基于信息瓶颈准则约束的对抗鲁棒语义通信方法,给出了语义通信系统模型, 并从互信息理论的角度分析了系统模型发射端、信道以及接收端中语义信息的任务相关和任务无关特征。在保留原始语义相似性的前提下,加入信息瓶颈准则约束,抑制解码器中间表征的任务无关特征,从而增强语义通信模型的抗干扰能力。通过实验以及综合分析,证明了该方法在提高基于深度学习的多级语义通信系统的对抗鲁棒性方面的优越性能。

Abstract:

Semantic communication based on deep learning aims to convey user intentions and semantic information, and is expected to become an important technical support for the “endogenous intelligence” architecture of 6G network. However, the adversarial robustness and security of semantic communication systems have not been fully studied. To this end, we proposed an adversarial robust semantic communication method based on information bottleneck criterion constraint. We firstly presented a semantic communication system model and then analyzed the task-relevant and task-irrelevant features of the semantic information in the transmitter, channel and receiver of the semantic communication system model from the perspective of mutual information theory. While preserving the original semantic similarity, we incorporated the information bottleneck criterion constraint to suppress the task-irrelevant features of the intermediate representation in the decoder, thereby enhancing the adversarial robustness of the semantic communication model. Through experiments and comprehensive analysis, we have demonstrated the superior performance of this method in improving the adversarial robustness of multi-level semantic communication systems based on deep learning.