Skip to content

A Frank-Wolfe Framework for Efficient and Effective Adversarial Attacks (AAAI'20)

License

Notifications You must be signed in to change notification settings

uclaml/Frank-Wolfe-AdvML

Repository files navigation

A Frank-Wolfe Framework for Efficient and Effective Adversarial Attacks (AAAI'20)

This is the code for AAAI'20 paper https://arxiv.org/abs/1811.10828 "A Frank-Wolfe Framework for Efficient and Effective Adversarial Attacks" by Jinghui Chen, Dongruo Zhou, Jinfeng Yi, and Quanquan Gu.

Prerequisites

  • Python (3.6.9)
  • Tensorflow (1.15.0)
  • Inception/ResNet pre-trained model
  • Download ImageNet validation set and put them in /imagenetdata/imgs/ folder

Usage Examples:

Pretrained Models:

  • Setup Inception V3 model:
  -  python3 setup_inception_v3.py
  • Setup ResNet model:
  -  python3 setup_resnet.py

Arguments:

  • arch: network architecture, e.g. "inception", "resnet"
  • sample: number of samples to attack
  • eps: epsilon, value 0.0 to enable grid search
  • att_iter: maximum number of iterations per attack
  • att_lr: attack learning rate (step size)
  • grad_est: zeroth-order gradient estimation batch size
  • sensing: type of sensing vectors, e.g. "gaussian", "sphere"
  • beta1: mementum parameter for FW
  • order: attack threat model type ("2" or "inf")

Basic Running Examples:

  • Run white-box attack on Inception V3 model:
  -  CUDA_VISIBLE_DEVICES=0 python3 test_attack.py --arch "inception" --method "FW" --order "inf" --sample 250 --eps 0.05 --att_lr 0.1 --beta1 0.9
  • Run black-box attack on ResNet V2 model:
  -  CUDA_VISIBLE_DEVICES=0 python3 test_attack_black.py --arch "resnet" --method "FW" --order "inf" --sample 1000 --eps 0.3 --att_lr 0.8 --grad_est 25 --sensing "sphere"

About

A Frank-Wolfe Framework for Efficient and Effective Adversarial Attacks (AAAI'20)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages