%0 Conference Paper %1 salman2019provably %A Salman, Hadi %A Yang, Greg %A Li, Jerry %A Zhang, Pengchuan %A Zhang, Huan %A Razenshteyn, Ilya %A Bubeck, Sebastien %D 2019 %K adversarial neurips2019 optimization robustness %T Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers %U http://arxiv.org/abs/1906.04584 %X Recent works have shown the effectiveness of randomized smoothing as a scalable technique for building neural network-based classifiers that are provably robust to $\ell_2$-norm adversarial perturbations. In this paper, we employ adversarial training to improve the performance of randomized smoothing. We design an adapted attack for smoothed classifiers, and we show how this attack can be used in an adversarial training setting to boost the provable robustness of smoothed classifiers. We demonstrate through extensive experimentation that our method consistently outperforms all existing provably $\ell_2$-robust classifiers by a significant margin on ImageNet and CIFAR-10, establishing the state-of-the-art for provable $\ell_2$-defenses. Moreover, we find that pre-training and semi-supervised learning boost adversarially trained smoothed classifiers even further. Our code and trained models are available at http://github.com/Hadisalman/smoothing-adversarial .

@inproceedings{salman2019provably, abstract = {Recent works have shown the effectiveness of randomized smoothing as a scalable technique for building neural network-based classifiers that are provably robust to $\ell_2$-norm adversarial perturbations. In this paper, we employ adversarial training to improve the performance of randomized smoothing. We design an adapted attack for smoothed classifiers, and we show how this attack can be used in an adversarial training setting to boost the provable robustness of smoothed classifiers. We demonstrate through extensive experimentation that our method consistently outperforms all existing provably $\ell_2$-robust classifiers by a significant margin on ImageNet and CIFAR-10, establishing the state-of-the-art for provable $\ell_2$-defenses. Moreover, we find that pre-training and semi-supervised learning boost adversarially trained smoothed classifiers even further. Our code and trained models are available at http://github.com/Hadisalman/smoothing-adversarial .}, added-at = {2019-12-10T01:06:35.000+0100}, author = {Salman, Hadi and Yang, Greg and Li, Jerry and Zhang, Pengchuan and Zhang, Huan and Razenshteyn, Ilya and Bubeck, Sebastien}, biburl = {https://www.bibsonomy.org/bibtex/23a7902c5c510a31ff2e3b3c0a29f38ef/kirk86}, description = {[1906.04584] Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers}, interhash = {6fa0ca602587c0575b3dfe9383a91278}, intrahash = {3a7902c5c510a31ff2e3b3c0a29f38ef}, keywords = {adversarial neurips2019 optimization robustness}, note = {cite arxiv:1906.04584Comment: Spotlight at the 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada; 9 pages main text; 31 pages total}, timestamp = {2019-12-10T01:06:35.000+0100}, title = {Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers}, url = {http://arxiv.org/abs/1906.04584}, year = 2019 }