Adversarial Example Resistant Hyperparameters and Deep Learning Networks

Abstract

Carefully crafted input has been shown to cause misclassifications in machine learning based classification systems resulting in the phenomenon of adversarial examples. Hyperparameters, the settings used to build and train machine learning models, have been shown to build machine learning models that are more resistant to adversarial examples. In this paper, we expand the research of hyperparameter saliency and incorporate deep learning architectures to compliment the field of research in addition to exploring the relationships between adversarial resistance and accuracy as well as depth. We find that hidden layer structures as well as activation function are important to resistance of adversarial perturbations, network depth provides for more robustness with some attacks while architecture influences robustness against others, and salient hyperparameter impact on accuracy is complex.