A Simulation Framework for Electromagnetic Signal Injection Attacks on Image Sensors

Abstract

Image sensors are fundamental to many intelligent systems, allowing visual perception and AI-driven decision-making. However, their integrity can be compromised by electromagnetic signal injection attacks (ESIA), which manipulate captured images without modifying sensor hardware or software. Despite the growing threat, system-level understanding of the attacks, as well as the development of defenses, remains limited, in part because collecting adversarial data is often complex and requires specialized attack setups. To address this challenge, we model ESIA and develop a simulation framework for generating synthetic adversarial images. Our analysis shows that these synthetic images are statistically indistinguishable from those produced by real attacks. The proposed framework enables faster vulnerability evaluation of computer vision (CV) algorithms, without the need for dedicated attack hardware. We also present a pilot study showing that the robustness of the algorithms can be improved by adversarial training, demonstrating a practical and scalable path toward mitigating ESIA threats.