Constraining neural networks for robustness through alternative encoding

Abstract

A neural network is augmented to enhance robustness against adversarial attack. In this approach, a fully-connected additional layer is associated with a last layer of the neural network. The additional layer has a lower dimensionality than at least one or more intermediate layers. After sizing the additional layer appropriately, a vector bit encoding is applied. The encoding comprises an encoding vector for each output class. Preferably, the encoding is an n-hot encoding, wherein n represents a hyperparameter. The resulting neural network is then trained to encourage the network to associated features with each of the hot positions. In this manner, the network learns a reduced feature set representing those features that contain a high amount of information with respect to each output class, and/or to learn constraints between those features and the output classes. The trained neural network is used to perform a classification that is robust against adversarial examples.