Manipulating Semantic Communication by Adding Adversarial Perturbations to Wireless Channel

To break through the transmission rate bottleneck of traditional communication, semantic communication is proposed to support emerging applications with extremely low latency requirements such as remote surgery and autonomous vehicle. Unlike the transmission of verbose symbols in traditional communication, mainstream semantic communications use deep learning technology to extract compact semantic information from data and convey it. However, the application of deep neural networks also poses security concerns, i.e., vulnerabilities to adversarial attacks. In this paper, we perform the first study on the security of semantic communication against both whitebox and black-box attacks by compromising the wireless channel between the transmitter and receiver. To launch practical and effective attacks, a systematic and universal attack framework is designed to craft content-agnostic, undetectable, robust whitebox perturbation signals as well as highly-transferable blackbox ones. Extensive experiments on two open-source datasets demonstrate that our attack framework can achieve over 87%, 99%, and 89% success rates in untargeted white-box, targeted white-box, and untargeted black-box attacks. This means that the proposed attack methods could severely threaten the quality of service of current semantic communications. We also propose two mitigation methods to resist such attacks.