First-order adversarial vulnerability of neural networks and input dimension: Code and Supplements

This repository contains code to accompany the ICML 2019 paper First-order adversarial vulnerability of neural networks and input dimension.

Packages

Special packages used:

• pytorch (tested on 1.0.0)
• foolbox (tested on 2.0.0)
• configargparse

Quick-start

• Train a net with default configuration and compute its vulnerability:

    python main.py

• Train a net on CIFAR10 with PGD and compute its vulnerability, with simililar settings than Section 4.1:

    python main.py --config configs/config_cifar_pgd.txt

• Train a net on upsampled CIFAR10 (bCIFAR10) and compute its vulnerability, with similar settings than Section 4.2:

    python main.py --config configs/config_bcifar.txt

• Any parameter from the config file can be overwritten from the command line, as in:

    python main.py --config configs/config_bcifar.txt --img_size 64 --log_step 10

See files in configs/ for other typical experiment settings/arguments.

Remark:

The experiments in Section 4.2 actually do not use the vulnerabilities contained in the vulnerability_XX.pt files (computed using the compute_vulnerability function from vulnerability.py). Instead, they use the test-set vulnerabilities contained in state['logs'] of the file last.pt (and computed using our custom pgd implementation from penalties.py), averaged over the 20 last training epochs.

Datasets

The code was implemented for 3 datasets:

• MNIST
• CIFAR10 (and upsampled CIFAR10, see paper Sec 4.2)
• a custom 12-class mini-ImageNet dataset (see below & paper Sec 4.2).

Getting our custom 12-class Mini-ImageNet Dataset

The custom mini-ImageNet was constructed by merging some similar classes together from the usual ImageNet dataset (for example all the dogs). If the resulting class got too big (e.g. the dog-class), we sampled a random subset of it. The 12 classes obtained were:

ball - car - cat - dog - elephant - fish - monkey - plane - ship - string-instrument - truck - wind-instrument

The list of images used in each class can be found in the imgnet12/ folder. To use our mini-imagenet dataloader, create a folder named imgnet12/train/ (resp. imgnet12/val/), with 12 subfolders (1 per class; name irrelevant) containing the images listed in train.txt (resp. val.txt).