Harnessing Edge Information for Improved Robustness in Vision Transformers
Authors: Yanxi Li, Chengbin Du, Chang Xu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments cover a wide range of robust benchmarks, including white-box and black-box adversarial attacks on Image Net-1K, employing FGSM (Goodfellow, Shlens, and Szegedy 2014) and PGD (Madry et al. 2017). |
| Researcher Affiliation | Academia | Yanxi Li, Chengbin Du, Chang Xu School of Computer Science, University of Sydney, Australia yali0722@uni.sydney.edu.au, chdu5632@uni.sydney.edu.au, c.xu@sydney.edu.au |
| Pseudocode | No | The paper contains architectural diagrams and mathematical equations but no pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any statement or link indicating the release of open-source code for the described methodology. |
| Open Datasets | Yes | Our experiments cover a wide range of robust benchmarks, including white-box and black-box adversarial attacks on Image Net-1K... In the domain of natural adversarial examples, we use the Image Net-A dataset (Hendrycks et al. 2021b). In the domain of out-of-distribution data, we use the Image Net-R dataset (Hendrycks et al. 2021a)... In the domain of common corruptions, we use the Image Net-C dataset (Hendrycks and Dietterich 2019)... |
| Dataset Splits | No | The paper mentions 'Training' and 'Evaluations' sections, and describes hyperparameters used for training, but does not explicitly detail a validation dataset split or its role in the experimental process. |
| Hardware Specification | Yes | We assess the throughput using a single NVIDIA RTX4090 GPU with 24GB of memory. |
| Software Dependencies | No | The paper mentions the use of certain algorithms and optimizers (e.g., "Canny edge detector", "SGD optimizer") and frameworks (e.g., Vision Transformers), but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | For the joint training objective in Eq. 9, we set the hyper-parameters α = 1.2 and β = 0.8. We use FGSM with l -norm for adversarial training and adopt a perturbation magnitude of ε = 1/255. We use the SGD optimizer, with a fixed learning rate of 1 10 4, a momentum of 0.9, and a weight decay of 2 10 5. For PGD, we execute it for 5 steps, using a step size of 0.5/255. |