On the Utility-Informativeness-Security Trade-Off in Discrete Task-Oriented Semantic Communication

Task-oriented semantic communication (ToSC) has been applied to edge inference tasks with limited communication and computing resources. By encoding the task-related features into a finite-size codebook and transmitting the index of the codebook, ToSC can be compatible with existing discrete communication systems. In such discrete ToSC systems, the codebook shared by the transmitter and receiver contains the information of original data and affects the task inference performance, which may also be obtained by adversaries. Thus, there exists an inherent utility-informativeness-security (UIS) trade-off problem in ToSC systems. This letter introduces a novel framework, named UIS-ToSC, which leverages the vector quantized variational information bottleneck (VQ-VIB) scheme for the trade-off issue. Furthermore, we exploit adversarial learning (AL) to train the system against information leakage. Comprehensive experiments demonstrate that the proposed scheme can efficiently reduce communication overhead and maintain information security with little influence on task-inference utility.