MultiGuard

Official code for paper MultiGuard: Provably Robust Multi-label Classification against Adversarial Examples, accepted by Neurips 2022.

Pretrained robust models will be released soon.

Dataset Preparation

Please place PASCAL-VOC in ./voc2007 , MS-COCO in COCO, NUS-WIDE in ./NUS_WIDE.

Evaluation

Place pretrained models under ./models

Evaluate perturbation range from [0,3], on VOC2007 dataset, using $k=5,k'=1,\sigma=0.5,\alpha=10^{-3}$, test on 500 images.

python3 eval.py --begin=0 --end=3 --T=300 --N=1000  --sigma=0.5 --batch_size=32 --k=5 --k_prime=1 --alpha=0.001 --record=test_voc_result.txt --dataset_type=PASCAL-VOC --model_name=tresnet_xl --model_path=./models/voc_asl_0.5.pth --M=500

Evaluate on MS-COCO dataset, using $k=10,k'=2,\sigma=0.5$.

python3 eval.py --begin=0 --end=3 --T=300 --N=1000  --sigma=0.5 --batch_size=32 --k=10 --k_prime=2 --alpha=0.001 --record=test_coco_result.txt --dataset_type=MS-COCO --model_name=tresnet_l --model_path=./models/coco_asl_0.5.pth --M=500

Evaluate on NUS-WIDE dataset, using $k=10,k'=2,\sigma=0.5$.

python3 eval.py --begin=0 --end=3 --T=300 --N=1000  --sigma=0.5 --batch_size=32 --k=10 --k_prime=2 --alpha=0.001 --record=test_nus_result.txt --dataset_type=NUS-WIDE --model_name=tresnet_l --model_path=./models/nus_asl_0.5.pth --M=500

Train

Train randomized smoothing model on VOC2007 dataset with $\sigma=0.5$.

python3 train.py --batch_size=16 --dataset_type=PASCAL-VOC --model_name=tresnet_xl --sigma=0.5 --log=train_voc.txt