Lightweight and Robust Wireless Semantic Communications
G. Chen, Guoshun Nan, Zhanghao Jiang, Hang Du, Ruisheng Shi, Qimei Cui, Xiaofeng Tao
Abstract
In this letter, we present SemGuarder, a novel deep learning-based semantic communication (DLSC) system that simultaneously incorporates physical-layer semantic encryption and adversarial identification, aiming to tackle multiple malicious attacks from open wireless channels. Compared to existing ones, our SemGuarder is more lightweight and more robust, facilitating the practical deployment of DLSC on mobile devices for security-critical applications. Specifically, SemGuarder consists of four key components including SemEncoder, SemDecoder, SemEncry, and SemDef. The former two modules comprise a novel lightweight block structure tailored for semantic representation learning. SemEncry employs physical-layer AES encryption against eavesdropping, while SemDef aims to detect and then mitigate adversarial attacks. Consequently, our SemGuarder is able to defend against various types of threats. We conduct experiments on two public benchmarks to validate the effectiveness of our SemGuarder for DLSC.