wzekai99/fat

Code for the paper "On the Tradeoff between Robustness and Fairness" (NeurIPS 2022)

On the Tradeoff between Robustness and Fairness

Code for NeurIPS'22 paper, "On the Tradeoff between Robustness and Fairness" by Xinsong Ma, Zekai Wang, Weiwei Liu.

Preferred Prerequisites

• Python (3.8)
• Pytorch (1.9.0)
• numpy (1.21)
• tensorboard (2.7)
• CUDA (11.1)

Simple Instructions

Train Resnet50 on CIFAR 100 using AT(Madry) with perturbation raius=0.015

CUDA_VISIBLE_DEVICES='0' python pgd_var.py --epsilon 0.015 --net resnet50 --dataset cifar100 --seed 1 --batch_size 256 --no_regular

Use our FAT on AT(Madry) with regularization = 0.1

CUDA_VISIBLE_DEVICES='0' python pgd_var.py --epsilon 0.015 --net resnet50 --dataset cifar100 --seed 1 --batch_size 256 --alpha 0.1

Train Resnet18 on CIFAR 10 using original TRADES with beta=6.0, perturbation raius=0.062

CUDA_VISIBLE_DEVICES='0' python trades_var.py --epsilon 0.062 --net resnet18 --beta 6.0 --dataset cifar10 --batch_size 256 --seed 1 --no_regular

Use our FAT on TRADES with regularization = 0.1

CUDA_VISIBLE_DEVICES='0' python trades_var.py --epsilon 0.062 --net resnet18 --beta 6.0 --dataset cifar10 --batch_size 256 --seed 1  --alpha 0.1

Evaluate models using FGSM, PGD and CW attack

CUDA_VISIBLE_DEVICES='0' python var_acc_test.py --attack_type fgsm --dataset 'cifar10' --dir_path './test_models'

CUDA_VISIBLE_DEVICES='0' python var_acc_test.py --attack_type pgd20 --dataset 'cifar10' --dir_path './test_models'

CUDA_VISIBLE_DEVICES='0' python var_acc_test.py --attack_type cw --dataset 'cifar10' --dir_path './test_models'

You can replace the device parameter 'CUDA_VISIBLE_DEVICES' and parameter '-- epsilon', '-- seed' if you need.

fat's Issues

The difference between Var and VCAR

Hello, it is my pleasure to read your paper and I have some questions I would like to inquire about, if it is not too much trouble.
First, I want to know the difference between Var and VCAR, and why not using Var as a regularization?
Second, about the proof of the Theorem 6.1, Jensen's inequality, as I know, is f(E(x)) <= E(f(x)), so I do not understand here.
I'm sorry to disturb you and I'm looking forward to your reply. Thanks