We propose a new white box adversarial attack method named EWR-PGD which exceeds the state-of-the-art attacks performance. It is more efficient than the state-of-the-art ODI-PGD method.
Code will be available soon.
When reducing the models to the same accuracy, the number of restarts required by the EWR-PGD significantly less than that of the ODI-PGD. EWR-PGD is up to roughly 5 times faster than ODI-PGD.
Figure 1. On 10 state-of-the-art defense models, comparison of the number of restarts required(the lower the better) when the EWR-PGD and ODI-PGD methods reduce the models to the same accuracy.
The models are available online:
- TRADES-Small Cnn, TRADES-WRN
- MadryLab-WRN
- MART-ResNet18, MART-WRN
- FBTF-ResNet18
- pytorch ResNeXt101-DenoiseAll, ResNet152-Denoise, ResNet152-Baseline
EWR-PGD ranks first on the TRADES white-box MNIST and CIFAR-10 leaderboards, reducing the accuracy of their MNIST model to 92.52% and the accuracy of their CIFAR-10 model to 52.95%. EWR-PGD also ranks first on MardyLab’s White-box CIFAR-10 leaderboard, reducing the accuracy of their CIFAR-10 model to 43.96%.
Table 1. Accuracy(the lower the better) under EWR-PGD and SOTA attacks and corresponding complexity.
dataset | model | EWR-PGD | EWR-PGD complexity | SOTA | SOTA complexity |
---|---|---|---|---|---|
MNIST | TRADES-SMN | 92.53%±0.01% | (20+300)×800 | 92.58% | ------------- |
CIFAR-10 | TRADES-WRN | 52.98%±0.02% | (5+100)×30 | 53.01% | (10+150)×20 |
CIFAR-10 | MadryLab-WRN | 43.98%±0.02% | (5+100)×30 | 43.99% | (10+150)×20 |
EWR-PGD ranks first among 1701 teams in CIKM2020 Analyticup: Alibaba-Tsinghua Adversarial Challenge on Object Detection. Surpassing the runner-up approach by∼14% in terms of scores.
Please contact liuye_ly94@163.com if you have any question.Enjoy!