ReadMe

Getting started

Let's start by installing all the dependencies.

pip3 install -r requirement.txt

Training

We employ cifar10_RT_ER.py and cifar10_JTDMoE.py to adversarially train an MoE model and a dual-model on the CIFAR-10 dataset using Resnet18. Similarly, we use tinyimagenet_RT_ER.py and tinyimagenet_JTDMoE.py for adversarial training of an MoE model and a dual-model on the TinyImageNet dataset using ViT-small. Below, we outline the key arguments and their usage.

• --net Specifies the model architecture used for training.
• --resume Defines the path to the checkpoint for evaluation or resuming training.
• --sizeSets the size of the input image.
• --betaControls the proportion of expert contributions in the model.
• --alphaAdjusts the contribution of the robust model within the dual-model.

Evaluation with AutoAttack

Please use the file AutoAttack.py to evaluate the model using AutoAttack.

Commands

To train an MoE on CIFAR-10 dataset using RT-ER method:

python3 cifar10_RT_ER.py --net res18_moe

To joint-training a dual-model on CIFAR-10 dataset using JTDMoE method:

python3 cifar10_JTDMoE.py --net dual_model_res18

To train an MoE on TinyImageNet dataset using RT-ER method:

python3 tinyimagenet_RT_ER.py --net vit_moe

To joint-training a dual-model on TinyImageNet dataset using JTDMoE method:

python3 tinyimagenet_JTDMoE.py --net dual_model_vit

Dataset Preparation

TinyImageNet

To obtain the original TinyImageNet dataset, please run the following scripts:

cd data/tiny_imagenet
wget http://cs231n.stanford.edu/tiny-imagenet-200.zip
unzip -qq 'tiny-imagenet-200.zip'
rm tiny-imagenet-200.zip