tuananhbui89 / tamoo

Task-Oriented Multi-Objective Optimization

Task-Oriented Multi-Objective Optimization

This repository contains the code for the paper: Generating Adversarial Examples with Task Oriented Multi-Objective Optimization, accepted to TMLR 2023. Link to the paper: https://openreview.net/forum?id=2f81Q622ww

Getting pre-trained models

To get pretrained models (i.e., with adversarial training), please run the script below:

    bash run_train.sh

For manual training, there are some parameters that need to be considered:

• arch: architecture code name which can be {resnet18, vgg16, efficientnet, googlenet, wideresnet}. Further information can be found in utils_arch.py.
• num_models: number of ensemble members, for examples, if arch==resnet18 and num_models==3 which refers to an ensemble of three resnet18 models. The default setting is num_models=1 which means that it is a single model.
• dataset: training dataset, either cifar10 or cifar100.
• method: type of training method (i.e., trainer which has been defined in trainer.py). For examples, ADV_Trainer refers to adversarial training method, which is the default setting.

Evaluating attacks

To get results of three main attack settings (ENS, EoT and UNI), please run the following scripts:

ENS - Generating Attack for Ensemble of Models.

    bash run_ensemble.sh

EoT - Generating Robust Attack against Ensemble of Transformations.

    bash run_eot.sh

UNI - Generating Universal perturbations.

    bash run_uniper.sh

To get results on evaluating the transferability of adverasrial examples we need to run in two phases. The first phase is to generate and save adversarial examples to a checkpoint file using the script below:

    bash run_genadv.sh

The second phase is to load the adversarial examples to attack other models using the script below:

    bash run_transferability.sh

To get results of adversarial training, please run the following script:

    bash run_adv_train.sh

To get results on attacking ImageNet dataset with Robustbench pretrained models, please run the following script:

    bash run_ens_imagenet.sh
    bash run_eot_imagenet.sh

Requirements

• advertorch==0.2.3
• robustbench

Citation

If you find this repository useful, please cite our paper:

@article{bui2023generating,
title={Generating Adversarial Examples with Task Oriented Multi-Objective Optimization},
author={Anh Tuan Bui and Trung Le and He Zhao and Quan Hung Tran and Paul Montague and Dinh Phung},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2023},
url={https://openreview.net/forum?id=2f81Q622ww},
note={}
}