Constant-Time Activation Functions Prevent Timing Side-Channel Leaks in Edge ML Inference

arXiv·low signal

Proposes and validates a constant-time implementation methodology for neural network activation functions (ReLU, sigmoid, tanh, GELU, Swish) on ARM Cortex-M4 microcontrollers, preventing timing side-channel information leakage during inference. Combines branchless selection, fixed-cost Pade-based approximation, dummy arithmetic, and cycle alignment. Relevant for teams deploying ML models on embedded devices in security-sensitive applications where inference timing can leak model architecture or input characteristics.