Multi-Modal Task-Oriented Secure Semantic Communication: A Hide-and-Deceive Approach

In this paper, we propose a multi-modal task-oriented secure semantic communication (SemCom) framework. In this framework, the base station (BS) jointly transmits the semantic information, i.e., semantic vectors of multi-modal data, to the user for multi-modal downstream task such as visual question answering (VQA). A powerful eavesdropper intends to complete the same multi-modal task by the eavesdropped semantic information and the stolen downstream task model. To degrade the task performance of the eavesdropper, an invertible neural network based hide-and-deceive (INN-HD) method is proposed. This method can deceive the eavesdropper by generating a dummy semantic information of one modality while hiding the original semantic information of that modality into another modality. However the user can extract the hidden semantic information by invertible neural network (INN) and complete task normally. Experimental results demonstrate that the proposed method can reduce the eavesdropper's downstream task performance by up to 100%, while maintaining 96% of the user's task performance compared to directly transmitting the original semantic vectors.