A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks

This project is for the paper "A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks". Some codes are from odin-pytorch, LID, and adversarial_image_defenses.

Preliminaries

It is tested under Ubuntu Linux 16.04.1 and Python 3.6 environment, and requries Pytorch package to be installed:

• Pytorch: Only GPU version is available.
• scipy
• scikit-learn

Downloading Out-of-Distribtion Datasets

We use download links of two out-of-distributin datasets from odin-pytorch:

• Tiny-ImageNet (resize)
• LSUN (resize)

Please place them to ./data/.

Downloading Pre-trained Models

We provide six pre-trained neural networks (1) three DenseNets trained on CIFAR-10, CIFAR-100 and SVHN, where models trained on CIFAR-10 and CIFAR-100 are from odin-pytorch, and (2) three ResNets trained on CIFAR-10, CIFAR-100 and SVHN.

• DenseNet on CIFAR-10 / DenseNet on CIFAR-100 / DenseNet on SVHN
• ResNet on CIFAR-10 / ResNet on CIFAR-100 / ResNet on SVHN

Please place them to ./pre_trained/.

Detecting Out-of-Distribution Samples (Baseline and ODIN)

# model: ResNet, in-distribution: CIFAR-10, gpu: 0
python OOD_Baseline_and_ODIN.py --dataset cifar10 --net_type resnet --gpu 0

Detecting Out-of-Distribution Samples (Mahalanobis detector)

1. Extract detection characteristics:

# model: ResNet, in-distribution: CIFAR-10, gpu: 0
python OOD_Generate_Mahalanobis.py --dataset cifar10 --net_type resnet --gpu 0

2. Train simple detectors:

# model: ResNet
python OOD_Regression_Mahalanobis.py --net_type resnet

Detecting Adversarial Samples (LID & Mahalanobis detector)

0. Generate adversarial samples:

# model: ResNet, in-distribution: CIFAR-10, adversarial attack: FGSM  gpu: 0
python ADV_Samples.py --dataset cifar10 --net_type resnet --adv_type FGSM --gpu 0

1. Extract detection characteristics:

# model: ResNet, in-distribution: CIFAR-10, adversarial attack: FGSM  gpu: 0
python ADV_Generate_LID_Mahalanobis.py --dataset cifar10 --net_type resnet --adv_type FGSM --gpu 0

2. Train simple detectors:

# model: ResNet
python ADV_Regression.py --net_type resnet