Better Diffusion Models Further Improve Adversarial Training

Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, Shuicheng Yan
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:36246-36263, 2023.

Abstract

It has been recognized that the data generated by the denoising diffusion probabilistic model (DDPM) improves adversarial training. After two years of rapid development in diffusion models, a question naturally arises: can better diffusion models further improve adversarial training? This paper gives an affirmative answer by employing the most recent diffusion model which has higher efficiency ($\sim 20$ sampling steps) and image quality (lower FID score) compared with DDPM. Our adversarially trained models achieve state-of-the-art performance on RobustBench using only generated data (no external datasets). Under the $\ell_\infty$-norm threat model with $\epsilon=8/255$, our models achieve $70.69\\%$ and $42.67\\%$ robust accuracy on CIFAR-10 and CIFAR-100, respectively, i.e. improving upon previous state-of-the-art models by $+4.58\\%$ and $+8.03\\%$. Under the $\ell_2$-norm threat model with $\epsilon=128/255$, our models achieve $84.86\\%$ on CIFAR-10 ($+4.44\\%$). These results also beat previous works that use external data. We also provide compelling results on the SVHN and TinyImageNet datasets. Our code is at https://github.com/wzekai99/DM-Improves-AT.

Cite this Paper


BibTeX
@InProceedings{pmlr-v202-wang23ad, title = {Better Diffusion Models Further Improve Adversarial Training}, author = {Wang, Zekai and Pang, Tianyu and Du, Chao and Lin, Min and Liu, Weiwei and Yan, Shuicheng}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {36246--36263}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/wang23ad/wang23ad.pdf}, url = {https://proceedings.mlr.press/v202/wang23ad.html}, abstract = {It has been recognized that the data generated by the denoising diffusion probabilistic model (DDPM) improves adversarial training. After two years of rapid development in diffusion models, a question naturally arises: can better diffusion models further improve adversarial training? This paper gives an affirmative answer by employing the most recent diffusion model which has higher efficiency ($\sim 20$ sampling steps) and image quality (lower FID score) compared with DDPM. Our adversarially trained models achieve state-of-the-art performance on RobustBench using only generated data (no external datasets). Under the $\ell_\infty$-norm threat model with $\epsilon=8/255$, our models achieve $70.69\\%$ and $42.67\\%$ robust accuracy on CIFAR-10 and CIFAR-100, respectively, i.e. improving upon previous state-of-the-art models by $+4.58\\%$ and $+8.03\\%$. Under the $\ell_2$-norm threat model with $\epsilon=128/255$, our models achieve $84.86\\%$ on CIFAR-10 ($+4.44\\%$). These results also beat previous works that use external data. We also provide compelling results on the SVHN and TinyImageNet datasets. Our code is at https://github.com/wzekai99/DM-Improves-AT.} }
Endnote
%0 Conference Paper %T Better Diffusion Models Further Improve Adversarial Training %A Zekai Wang %A Tianyu Pang %A Chao Du %A Min Lin %A Weiwei Liu %A Shuicheng Yan %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-wang23ad %I PMLR %P 36246--36263 %U https://proceedings.mlr.press/v202/wang23ad.html %V 202 %X It has been recognized that the data generated by the denoising diffusion probabilistic model (DDPM) improves adversarial training. After two years of rapid development in diffusion models, a question naturally arises: can better diffusion models further improve adversarial training? This paper gives an affirmative answer by employing the most recent diffusion model which has higher efficiency ($\sim 20$ sampling steps) and image quality (lower FID score) compared with DDPM. Our adversarially trained models achieve state-of-the-art performance on RobustBench using only generated data (no external datasets). Under the $\ell_\infty$-norm threat model with $\epsilon=8/255$, our models achieve $70.69\\%$ and $42.67\\%$ robust accuracy on CIFAR-10 and CIFAR-100, respectively, i.e. improving upon previous state-of-the-art models by $+4.58\\%$ and $+8.03\\%$. Under the $\ell_2$-norm threat model with $\epsilon=128/255$, our models achieve $84.86\\%$ on CIFAR-10 ($+4.44\\%$). These results also beat previous works that use external data. We also provide compelling results on the SVHN and TinyImageNet datasets. Our code is at https://github.com/wzekai99/DM-Improves-AT.
APA
Wang, Z., Pang, T., Du, C., Lin, M., Liu, W. & Yan, S.. (2023). Better Diffusion Models Further Improve Adversarial Training. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:36246-36263 Available from https://proceedings.mlr.press/v202/wang23ad.html.

Related Material