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

Despite their tremendous successes, convolutional neural networks (CNNs)\nincur high computational/storage costs and are vulnerable to adversarial\nperturbations. Recent works on robust model compression address these\nchallenges by combining model compression techniques with adversarial training.\nBut these methods are unable to improve throughput (frames-per-second) on\nreal-life hardware while simultaneously preserving robustness to adversarial\nperturbations. To overcome this problem, we propose the method of Generalized\nDepthwise-Separable (GDWS) convolution -- an efficient, universal,\npost-training approximation of a standard 2D convolution. GDWS dramatically\nimproves the throughput of a standard pre-trained network on real-life hardware\nwhile preserving its robustness. Lastly, GDWS is scalable to large problem\nsizes since it operates on pre-trained models and doesn't require any\nadditional training. We establish the optimality of GDWS as a 2D convolution\napproximator and present exact algorithms for constructing optimal GDWS\nconvolutions under complexity and error constraints. We demonstrate the\neffectiveness of GDWS via extensive experiments on CIFAR-10, SVHN, and ImageNet\ndatasets. Our code can be found at https://github.com/hsndbk4/GDWS.\n