Voltage Scaling-Agnostic Counteraction of Side-Channel Neural Net Reverse Engineering via Machine Learning Compensation and Multi-Level Shuffling

This work proposes a voltage scaling-agnostic counteraction against neural network weight reverse engineering via side-channel attacks. Multi-level shuffling and machine learning-based dual power compensation are introduced. State-of-the-art protection ($\gt200\cdot 10^{6}$ MTD) is achieved at low power overhead (1.76$\times $) and zero latency overhead.