qilong-zhang / cvpr2021-competition-unrestricted-adversarial-attacks-on-imagenet

Our Team (green hand) 6th Solution for CVPR-2021 AIC-VI: Unrestricted Adversarial Attacks on ImageNet

Our Team (green hand) 6th Solution for CVPR-2021 AIC-VI: Unrestricted Adversarial Attacks on ImageNet

• Our method (Transfer-based attacks):

  • Dong et al. (TI-BIM) [1]
  • Zou et al. (DEM) [3], i.e., the variant of Xie et al. (DI-FGSM) [2]
  • Wang et al. (Pre-gradient I-FGSM) [4]
  • Ours works
    • PI-FGSM [5] & Our official repo
    • PI-FGSM++ [6] & Our official repo
    • 🚀Staircase sign method (SSM) [7] & Our official repo
      • A simple test: By replacing the Sign method of DI-TI-FGSM with our SSM , we significantly increase the score from ~8 to ~22 (full score is 100)
  • Pre-process tricks:
    • smooth the images firstly
    • resize to 256 instead of 500 before fed into DNNs at each iteration

• Setting

  • 
  • Iteration T=40
  • Step size: ~ 0.6

• Substitute models:

  • DenseNet-121, DenseNet-169, ResNet-50, ResNet-101, VGG-19, EfficientNet-b5

• The cost of GPU memory: 12G on Titan xp (12G)

Implementation

• Requirement

  • Python 3.7
  • Pytorch 1.8.0
  • torchvision 0.9.0
  • pandas 1.1.3
  • matplotlib 3.3.4
  • scipy 1.5.4
  • timm 0.4.5
  • tqdm 4.43.0

• Download the dataset from here (or select one of the following datasets)

  • Round 1: https://tianchi-competition.oss-cn-hangzhou.aliyuncs.com/531853/imagenet_round1_210122.zip
  • Round 2: https://tianchi-competition.oss-cn-hangzhou.aliyuncs.com/531853/imagenet_round2_210325.zip

• Unzip the downloaded dataset zip, and put 5K images into "input_dir/images/" file.

• Then run the code

  python run.py

• Finally, the adversarial examples will be saved at "output_dir/"

Result

Reference

[1] Yinpeng Dong, Tianyu Pang, Hang Su and Jun Zhu: Evading defenses to transferable adversarial examples by translation-invariant attacks, CVPR 2019

[2] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren and Alan L Yuille: Improving Transferability of Adversarial Examples with Input Diversity, CVPR 2019

[3] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui and Wei Lin: Improving the Transferability of Adversarial Examples with Resized-Diverse-Inputs,Diversity-Ensemble and Region Fitting, ECCV 2020

[4] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang and Kun He: Boosting Adversarial Transferability through Enhanced Momentum, ArXiv 2021

[5] Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu and Heng Tao Shen: Patch-wise attack for fooling deep neural network, ECCV 2020

[6] Lianli Gao, Qilong Zhang, Jingkuan Song and Heng Tao Shen: Patch-wise++ Perturbation for Adversarial Targeted Attacks, ArXiv 2020

[7] Lianli Gao, Qilong Zhang, Xiaosu Zhu, Jingkuan Song and Heng Tao Shen: Staircase Sign Method for Boosting Adversarial Attacks, ArXiv 2021