Stealthy Backdoor Attacks on Semantic Symbols in Semantic Communications
Yuan Zhou, Rose Qingyang Hu, Yi Qian · 2024
Semantic communication is of crucial importance for the next-generation wireless communication networks. Recent advancements have primarily benefited from the design of semantic communication systems based on deep learning. Nevertheless, these deep learning-based systems are vulnerable to certain security attacks, particularly backdoor attacks. A novel attack paradigm, backdoor attacks on semantic symbols (BASS), targets reconstruction tasks by manipulating the reconstructed source data or features. However, the perceivable risks associated with BASS have not been thoroughly explored. This paper investigates the perceivable risks of BASS in the context of computer vision tasks. A transform-based methodology is designed to improve the stealthiness of the poisoned reconstructed target samples in the training dataset. In addition, while various hidden triggers have been studied for traditional backdoor attacks, they cannot be applied to BASS directly due to the unaligned model problem. To address this, an iterative hidden trigger generation (IHTG) algorithm is proposed. The simulation results demonstrate the effectiveness of the proposed methods in addressing the perceivable risks in BASS.