Formulating Robustness Against Unforeseen Attacks

Authors: Sihui Dai, Saeed Mahloujifar, Prateek Mittal

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate that using VR can lead to improved generalization to unforeseen attacks during test-time, and combining VR with perceptual adversarial training (Laidlaw et al., 2021) achieves state-of-the-art robustness on unforeseen attacks.
Researcher Affiliation Academia Sihui Dai Princeton University sihuid@princeton.edu Saeed Mahloujifar Princeton University sfar@princeton.edu Prateek Mittal Princeton University pmittal@princeton.edu
Pseudocode No The paper describes the training objective with equations and text, but does not provide any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is publicly available at https://github.com/inspire-group/variation-regularization.
Open Datasets Yes Datasets We train models on CIFAR-10, CIFAR-100, (Krizhevsky et al., 2009) and Image Nette (Howard).
Dataset Splits No The paper mentions training on datasets like CIFAR-10 and CIFAR-100 but does not explicitly provide the specific training, validation, and test split percentages or sample counts in the main text or appendix. While it refers to 'standard practices', explicit split details are not stated.
Hardware Specification Yes Training an AT-VR ResNet-18 on CIFAR-10 takes 12-15 hours on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions implementing models 'using PyTorch' but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes For CIFAR-10 and CIFAR-100, we used an ℓ adversarial budget of 8/255 and ℓ2 budget of 0.5. For Image Nette, we use an ℓ adversarial budget of 4/255 and ℓ2 budget of 1.0. We use a batch size of 128. For CIFAR-10 and CIFAR-100, we use an initial learning rate of 0.1 with a cosine learning rate schedule that decays to 0.001 at the last epoch. For Image Nette, we use an initial learning rate of 0.01 with cosine learning rate schedule that decays to 0.001. We train for 100 epochs on all datasets. We use a momentum of 0.9 and a weight decay of 0.0005. We use PGD to generate adversarial examples with 10 steps and step size of 2/255 for ℓ and 0.01 for ℓ2.