ExposureExperiment

Datasets

---

	Cifar10	Cifar100	Mnist	F-Mnist	MVTEC	Medical	SVHN
eps	8/255	8/255	8/255	8/255	2/255	2/255	8/255
size	32x32	32x32	32x32	32x32	224x224	224x224	32x32

Commands

---

python train_and_evaluate.py \
        --source_dataset $SOURCE_DATASET \
        --source_class $CLASS \
        --exposure_dataset $EXPOSURE_DATASET \
        --test_attacks ${TEST_ATTACKS[@]} \
        --batch_size $BATCH_SIZE \
        --max_epochs $MAX_EPOCHS \
        --train_attack_step $TRAIN_ATTACK_STEP \
        --attack_eps $ATTACK_EPS \
        --test_step $TEST_STEP \
        --save_step $SAVE_STEP \
        --cuda_device $CUDA_DEVICE \
        --loss_threshold $LOSS_THRESHOLD \
        --model $MODEL_ARCHITECTURE \
        --checkpoints_path $CHECKPOINT_PATH \
        --output_path $OUTPUT_PATH \
        --tensorboard_path $TENSORBOARD_PATH \
        --clean \
        --force_restart

More details on parameters

--source_dataset: Specifies the name of the normal dataset. Valid options are 'cifar10', 'cifar100', 'mnist', 'fashion', and 'svhn'.

--source_class: Specifies the index of the normal class in the normal dataset.

--exposure_dataset: Specifies the name of the exposure dataset. Valid options are 'cifar10', 'cifar100', 'mnist', 'fashion', and 'svhn'.

--checkpoints_path: Specifies the path to save the model checkpoints. The default path is './Model-Checkpoints/'.

--output_path: Specifies the path to which plots, results, and other data will be recorded. The default path is './results/'.

--tensorboard_path: Specifies the path to which plots, results, and other data will be recorded. The default path is './tensorboard/'.

--max_epochs: Specifies the maximum number of epochs to train the model. The default value is 30.

--batch_size: Specifies the size of each batch input to the model. The default value is 128.

--attack_eps: Specifies the attack eps used for both training and testing. The default value is 8/255.

--test_attacks: Specifies the desired attacks for adversarial testing. For example, '--test_attacks FGSM PGD-10 PGD-100'.

--train_attack_step: Specifies the desired attack step for adversarial training. The default value is 10.

--clean: Specifies whether to perform clean training (if present) or adversarial training

--test_step: Specifies the frequency at which to run tests. The default value is 1.

--save_step: Specifies the frequency at which to save model checkpoints. The default value is 1.

--cuda_device: Specifies the index of your CUDA device. The default value is 0.

--force_restart: Specifies whether to start training from scratch (if present) or use already available checkpoints

--loss_threshold: Specifies the loss threshold used for early stopping. The default value is 0.001.

--model: Specifies the backbone used for training. Valid options are 'preactresnet18', 'preactresnet34', 'preactresnet50', 'preactresnet101', and 'preactresnet152.

Outputs / Run

---

• Clean AUROC after normal training
• Clean AUROC after adversarial training
• Robust AUROC after adversarial training
• FID Score Support (in this repos)

TODO

---

• MVTEC Ram Efficient support