Jakub Breier

A single electromagnetic pulse can induce undetectable, persistent accuracy collapse in the Intel NCS2, even when the device is idle, revealing a critical vulnerability in edge AI inference.

Securing embedded neural networks against timing attacks is now more practical: constant-time activation functions can be implemented with high accuracy and low overhead.

TVLA misses subtle side-channel leakage in neural networks, but a new statistical test closes the gap.

Floating-point precision in embedded neural networks makes them shockingly vulnerable: a single electromagnetic fault injection can almost completely degrade accuracy, while integer representations offer significantly better resilience.