Abstract
The confidentiality of trained AI models on edge devices is at risk from side-channel attacks exploiting power and electromagnetic emissions. This paper proposes a novel training methodology to enhance resilience against such threats by introducing randomized and interchangeable model configurations during inference. Experimental results on Google Coral Edge TPU show a reduction in side-channel leakage and a slower increase in t-scores over 20,000 traces, demonstrating robustness against adversarial observations. The defense maintains high accuracy, with about 1% degradation in most configurations, and requires no additional hardware or software changes, making it the only applicable solution for existing Edge TPUs.