Generalized Depthwise-Separable Convolutions for Adversarially Robust and Efficient Neural Networks

But these methods are unable to improve throughput (frames-per-second) on real-life hardware while simultaneously preserving robustness to adversarial perturbations.