Synchronization-Free Backdoor Attacks on Edge-Deployed Neural Network for IoT Systems

Abstract

Deep Neural Networks integrated into sensor-driven embedded systems have become fundamental to edge computing applications. While Inertial Measurement Units (IMUs) are ubiquitous in these architectures due to their cost-efficiency, the reliance on data-driven models expands the attack surface of the embedded computing pipeline. This paper investigates the vulnerability of such systems by presenting a synchronization-free backdoor attack framework. We address two critical threat models in the system lifecycle: training-data poisoning and outsourced training. For the poisoning setting, we exploit the physical-digital signal interface by developing a frequency-domain injection method that utilizes sidelobe perturbations to implant stealthy triggers. For the outsourced scenario, we propose a dual-optimization joint trigger learning scheme that co-optimizes trigger patterns and model objectives, eliminating the dependency on synchronization with the victim’s training process. To ensure attack stealthiness within the operational bounds, a dynamic weighting strategy is introduced to balance attack efficacy with benign inference accuracy. Extensive experiments on standard benchmarks demonstrate that current embedded architectures are highly susceptible to these methods: the frequency-domain poisoning achieves over 60% average attack success, while the outsourced-training attack exceeds 95%, highlighting a critical security gap in sensor-based intelligent systems.