Adversarial Attacks on Time Series

Time series classification models have been garnering significant importance in the research community. However, not much has done on generating adversarial samples for these models. These can become a security concern. In this paper, we propose utilizing an transformation network (ATN) distilled model to attack various time The proposed utilizes as surrogate that mimics behavior of attacked classical Our methodology is applied onto 1-nearest neighbor dynamic warping (1-NN DTW) and fully convolutional (FCN), all which are trained 42 University California Riverside (UCR) datasets. show both were susceptible attacks When compared Fast Gradient Sign Method, generates larger faction successful black-box attacks. A simple defense mechanism successfully devised reduce fraction samples. Finally, recommend future researchers develop incorporating data into their training sets improve resilience