Deep neural networks are capable of training fast and generalizing well\nwithin many domains. Despite their promising performance, deep networks have\nshown sensitivities to perturbations of their inputs (e.g., adversarial\nexamples) and their learned feature representations are often difficult to\ninterpret, raising concerns about their true capability and trustworthiness.\nRecent work in adversarial training, a form of robust optimization in which the\nmodel is optimized against adversarial examples, demonstrates the ability to\nimprove performance sensitivities to perturbations and yield feature\nrepresentations that are more interpretable. Adversarial training, however,\ncomes with an increased computational cost over that of standard (i.e.,\nnonrobust) training, rendering it impractical for use in large-scale problems.\nRecent work suggests that a fast approximation to adversarial training shows\npromise for reducing training time and maintaining robustness in the presence\nof perturbations bounded by the infinity norm. In this work, we demonstrate\nthat this approach extends to the Euclidean norm and preserves the\nhuman-aligned feature representations that are common for robust models.\nAdditionally, we show that using a distributed training scheme can further\nreduce the time to train robust deep networks. Fast adversarial training is a\npromising approach that will provide increased security and explainability in\nmachine learning applications for which robust optimization was previously\nthought to be impractical.\n