Deterministic and Statistical Strategies to Protect ANNs against Fault Injection Attacks

Attificial neural networks are currently used for many tasks, including safety critical ones such as automated driving. Hence, it is very important to protect them against faults and fault attacks. In this work, we propose two fault injection attack detection mechanisms: one based on using output labels for a reference input, and the other on the activations of neurons. First, we calibrate our detectors during normal conditions. Thereafter, we verify them to maximize fault detection performance. To prove the effectiveness of our solution, we consider highly employed neural networks (AlexNet, GoogleNet, and VGG) with their associated dataset ImageNet. Our results show that for both detectors we are able to obtain a high rate of coverage against faults, typically above 96%. Moreover, the hardware and software implementations of our detector indicate an extremely low area and time overhead.