Data Augmentation Can Improve Robustness

Authors: Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, Timothy A Mann

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on CIFAR-10 against ℓ and ℓ2 norm-bounded perturbations of size ϵ = 8/255 and ϵ = 128/255, respectively. We show large absolute improvements of +2.93% and +2.16% in robust accuracy compared to previous state-of-the-art methods. We conduct thorough experiments to show that our approach generalizes across architectures, datasets and threat models.
Researcher Affiliation Industry Sylvestre-Alvise Rebuffi*, Sven Gowal*, Dan Calian, Florian Stimberg, Olivia Wiles and Timothy Mann Deep Mind, London {sylvestre,sgowal}@deepmind.com
Pseudocode No No explicit pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes The code written in JAX [4] and Haiku [26] is available online at https://github.com/ deepmind/deepmind-research/tree/master/adversarial_robustness.
Open Datasets Yes We evaluate our approach on CIFAR-10 against ℓ and ℓ2 norm-bounded perturbations of size ϵ = 8/255 and ϵ = 128/255, respectively. We also achieve a significant performance boost with this approach while using other architectures and datasets such as CIFAR-100, SVHN and TINYIMAGENET.
Dataset Splits Yes Specifically, we train two (and only two) models for each hyperparameter setting, perform early stopping for each model on a separate validation set of 1024 samples using PGD40 similarly to Rice et al. [44] and pick the best model by evaluating the robust accuracy on the same validation set .
Hardware Specification Yes We train for 400 epochs with a batch size of 512 split over 32 Google Cloud TPU v3 cores [4], and the learning rate is initially set to 0.1 and decayed by a factor 10 two-thirds-of-the-way through training.
Software Dependencies No The paper mentions JAX [4] and Haiku [26] but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes We train for 400 epochs with a batch size of 512 split over 32 Google Cloud TPU v3 cores [4], and the learning rate is initially set to 0.1 and decayed by a factor 10 two-thirds-of-the-way through training. We scale the learning rates using the linear scaling rule of Goyal et al. [21] (i.e., effective LR = max(LR batch size/256, LR)). The decay rate of WA is set to τ = 0.999.