Identifying artificial artifacts in input data to detect adversarial attacks

Abstract

Mechanisms are provided for training a classifier to identify adversarial input data. A neural network processes original input data representing a plurality of non-adversarial original input data and mean output learning logic determines a mean response for each intermediate layer of the neural network based on results of processing the original input data. The neural network processes adversarial input data and layer-wise comparison logic compares, for each intermediate layer of the neural network, a response generated by the intermediate layer based on processing the adversarial input data, to the mean response associated with the intermediate layer, to thereby generate a distance metric for the intermediate layer. The layer-wise comparison logic generates a vector output based on the distance metrics that is used to train a classifier to identify adversarial input data based on responses generated by intermediate layers of the neural network.