At this year’s International Conference on Learning Representations (ICLR), a group of researchers from the University of Maryland presented an attack technique meant to slow down deep learning models that have been optimized for fast and sensitive operations.