Mitigating Adversarial Effects Through Randomization
This paper proposed to utilize randomization to mitigate adversarial effects (https://arxiv.org/pdf/1711.01991.pdf). By combining the proposed randomization method with an adversarially trained model, it ranked No.2 among 107 defense teams in the NIPS 2017 adversarial examples defense challenge (https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack).
The approach
The main ideal of the defense is to utilize randomization to defend adversarial examples:
- Random Resizing: after pre-processing, resize the original image (size of 299 x 299 x 3) to a larger size, Rnd x Rnd x 3, randomly, where Rnd is within the range [310, 331).
- Random Padding: after resizing, pad the resized image to a new image with size 331 x 331 x 3, where the padding size at left, right, upper, bottom are [a, 331-Rnd-a, b, 331-Rnd-b]. The possible padding pattern for the size Rnd is (331-Rnd+1)^2.
In general, the pipeline is shown below:
Pros
- No additional training/finetuning is required
- Very little computation introduced
- Compatiable to different networks and different defending methods (i.e., we use randomization + ensemble adversarial training + Inception-Resnet-v2 in our submission)
Ensemble adversarial training model
Team Member
- Cihang Xie (Johns Hopkins University)
- Zhishuai Zhang (Johns Hopkins University)
- Jianyu Wang (Baidu Research)
- Zhou Ren (Snap Inc.)
Leaderboard
Our team name is iyswim, and our rank is No.2.
Citing this work
If you find this work is useful in your research, please consider citing:
@inproceedings{xie2017mitigating,
title={Mitigating Adversarial Effects Through Randomization},
author={Xie, Cihang and Wang, Jianyu and Zhang, Zhishuai and Ren, Zhou and Yuille, Alan},
booktitle={International Conference on Learning Representations},
year={2018}
}