Is Semantic Communication for Autonomous Driving Secured against Adversarial Attacks?
Soheyb Ribouh, Abdenour Hadid · 2024
Artificial intelligence (AI) has emerged as a key driver toward the development of the next generation of wireless communication systems (6G), where deep learning (DL) has been widely used for designing the semantic communication modules. However, the utilization of deep neural networks (DNN) may increase the vulnerability of the semantic communication to adversarial attacks. In this paper, we introduce a state-of-the-art wireless semantic communication system for vision-based autonomous driving, integrating a semantic encoder/decoder for image transmission designed based on DNN architecture. We thoroughly evaluate the vulnerability of the model to various adversarial attacks. An adversarial attack consists of slightly modifying the transmitted images by adding small perturbations perhaps imperceptible to humans but harmful to the model. We inject various adversarial attacks including: Auto-PGD attack, FSGM attack, DeepFool attack, Square attack, and adversarial patch attack. Evaluated on semantic communication tasks, the experiments clearly show that semantic communication is vulnerable to adversarial attacks. This urgently calls for defense mechanisms to ensure reliability and safety of wireless semantic communication.