A Learning-based Joint Source-Channel Coding Approach for Wireless Covert Image Transmission

Covert communications provide critical security to guarantee undetectable data transmission against the passive wardens. However, it is always a difficult problem to characterize the invisibility of covert communication, and conventional works usually leverage the theoretic bound of the power constraint on transmitted signal to achieve invisible transmission, which is unnecessary and inefficient in general. In this paper, we propose a learning-based joint source-channel coding (JSCC) approach for covert image transmission. The proposed method introduces a discriminator to evaluate transmission covertness, and adds its output as a regularization term to the loss function for training the neural network(NN)-based coders. This method enables the JSCC encoder to adjust its parameters for not only the quality of data recovery but also covert communication. The experimental results demonstrate that the proposed framework outperforms the conventional methods in both the data recovery performance and the transmission invisibility performance.