An Energy-Efficient Neural Network Accelerator with Improved Resilience Against Fault Attacks

Embedded neural network (NN) implementations are susceptible to misclassification under fault attacks. Injecting strong electromagnetic (EM) pulses is a non-invasive yet detrimental attack that affects the NN operations by (i) causing faults in the NN model/inputs while being read by the NN computation unit, and (ii) corrupting NN computations results to cause misclassification eventually. This paper presents the first ASIC demonstration of an energy-efficient NN accelerator with inbuilt fault attack detection. We incorporated lightweight cryptography-aided checks using the Craft cipher for on-chip verification to detect model/input errors and also as a fault detection sensor. Our developed ASIC has demonstrated excellent error detection capabilities (100% detection for 100k error attempts) with a minimal area overhead of 5.9% and negligible NN accuracy degradation.