Taxonomy Driven Fast Adversarial Training

Code for "Taxonomy Driven Fast Adversarial Training" (AAAI 2024)

Environments

python 3.7.13
torch 1.13.1
torchvision 0.13

Files

TDAT.py

Train ResNet18 on CIFAR-10 with our proposed TDAT.

python TDAT.py --batch-m 0.75 --delta-init "random" --out-dir "TDAT" --log "CIFAR10.log" --model "ResNet18" --lamda 0.6 --inner-gamma 0.15 --outer-gamma 0.15 --save-epoch 1

models

This folder holds the codes for backbones.

CIFAR10/CIFAR100/TinyImageNet/ImageNet100

These folders store the corresponding training log.

Trained Models

Checkpoint on CIFAR-10 with our method

bookman233 / TDAT