Data augmentation alone can improve adversarial training
Authors: Lin Li, Michael W. Spratling
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, our augmentation method achieves the state-of-the-art accuracy and robustness in adversarial training. ... Models were trained with each transformation at each hardness. ... Robustness is evaluated against PGD50. ... Tab. 1 it can be seen that our method, IDBH, achieves the state-of-the-art best and end robustness for data augmentations... |
| Researcher Affiliation | Academia | Lin Li & Michael Spratling Department of Informatics King s College London 30 Aldwych, London, WC2B 4BG, UK {lin.3.li, michael.spratling}@kcl.ac.uk |
| Pseudocode | Yes | Cropshift (Fig. 4; Algorithm 1) first randomly crops a region in the image and then shifts it around to a random location in the input space. ... Pseudo-code for the proposed augmentation procedure can be found in Appendix E. ... These two algorithms Cropshift and IDBH are illustrated in pseudo codes in Algorithm 1 and Algorithm 2 respectively. |
| Open Source Code | Yes | Code is available at: https://github.com/Tree LLi/DA-Alone-Improves-AT. ... To further facilitate the reproducibility, we are going to share our code and the pre-trained models with the reviewers and area chairs once the discussion forum is open, and will publish them alongside the paper if accepted. |
| Open Datasets | Yes | The dataset was CIFAR10 (Krizhevsky, 2009). ... We additionally evaluated on the datasets SVHN (Netzer et al., 2011) and Tiny Image Net (TIN) (Le & Yang, 2015). |
| Dataset Splits | Yes | To alleviate overfitting, Rice et al. (2020) propose to track the model s robustness on a reserved validation data and select the checkpoint with the best validation robustness instead of the one at the end of training. ... Models were trained by stochastic gradient descent for 200 epochs with initial learning rate 0.1, divided by 10 at the epochs 100 and 150. |
| Hardware Specification | Yes | Experiments were run on Tesla V100 and A100 GPUs. |
| Software Dependencies | Yes | Our methods including both Cropshift and IDBH can be easily implemented using the popular machine learning development frameworks like Py Torch (Paszke et al., 2019). ... Robustness was evaluated against AA using the implementation of Kim (2021). |
| Experiment Setup | Yes | Models were trained by stochastic gradient descent for 200 epochs with initial learning rate 0.1, divided by 10 at the epochs 100 and 150. The momentum was 0.9, the weight decay was 5e-4 and the batch size was 128. ... The full parameters of the optimal augmentation schedules we found are described in Tab. 12 and Tab. 11. The training and evaluation settings are described in Section 5 and Appendix C. |