Robust Task-Oriented Communication with Semantic-Aware Masking and Discrete Codebook

Task-oriented semantic communication has gained notable interest for its capacity to minimize transmitted data volume without sacrificing task performance. Previous research has mainly concentrated on random masking, which may obscure critical features and hinder the model's ability to learn transferable representations. In this paper, a robust semantic communication system based on semantic-aware masking and discrete codebook (SAMDC) is proposed. Specifically, we develop a semantic-aware sampling strategy, which can selectively mask image patches with low semantic importance instead of random masking, to enhance the model's capacity to learn semantic information and boost training efficiency. Moreover, we also apply an improved robust discrete codebook, shared between the transmitter and receiver. This codebook comprises orthogonal and trainable basis vectors that symbolize the encoded features, thereby enhancing the system's robustness. Experimental results demonstrate that our proposed robust SAMDC significantly enhances the processing efficiency of semantic information. This improvement leads to better performance in communication tasks, particularly in challenging low signal-to-noise ratio (SNR) situations.