GLRP: Guided by Layer-wise Relevance Propagation – Selecting Crucial Neurons in Artificial Neural Networks

Artificial neural networks (ANN) are used in critical application domains like autonomous driving or medicine. As a result of an attack or due to faults in the hardware, an ANN might make wrong predictions which might lead to a dangerous malfunction.Using layer-wise relevance propagation, we identify crucial neurons in trained ANNs. Hardening techniques, independent of our approach, can then be used to protect crucial neurons. Mathematically as well as with empirical experiments, we show that hardening the crucial neurons reduces or even eliminates the number of wrong predictions made by the ANN.