Multi-Agent Reinforcement Learning for Covert Semantic Communications over Wireless Networks

In this paper, a covert semantic communication framework is proposed for image transmission over wireless networks. In the proposed framework, devices extract and selectively transmit semantic information of image data to a base station (BS). The semantic information consists of the objects in the image and a set of attributes of each object. A warden selects a device to detect and eavesdrops the semantic information. To ensure the security of semantic communications, a jammer, acts as the defender, requires to find a vulnerable device and transmits jamming signals to the vulnerable device. The metric to measure the performance of the covert semantic communications is defined as the difference in the average accuracy of the BS and the warden answering a set of questions for each image. To maximize the performance of covert semantic communications, each device and the jammer must jointly optimize their transmit power, determine the vulnerable device to be protected, and determine the partial semantic information that each device needs to transmit. To solve this problem, we propose a multi-agent policy gradient (MAPG) algorithm. The proposed algorithm enables each device and the jammer to cooperatively discover the vulnerable devices as well as find the semantic information transmission and power control policies that maximize the performance of the covert semantic communication system. Simulation results show that the proposed algorithm can improve the communication performance by up to 14.5% compared to the independent reinforcement learning.