3SAT: A Simple Self-Supervised Adversarial Training Framework
Authors: Jiang Fang, Haonan He, Jiyan Sun, Jiadong Fu, Zhaorui Guo, Yinlong Liu, Wei Ma
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that 3SAT surpasses the known SOTA self-AT methods across all evaluation metrics on various datasets. Notably, on CIFAR-10, 3SAT improves the robust accuracy of the sota self-AT method by 16.19% and the standard accuracy by 11.41%. |
| Researcher Affiliation | Academia | 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China EMAIL |
| Pseudocode | No | The paper describes methods using narrative text and mathematical equations (e.g., Equation 1, 2, 3, 4, 5, 6) but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is at https://github.com/Meng Nan Fang/3SAT |
| Open Datasets | Yes | We evaluate the representation performance and robustness of 3SAT on different datasets: CIFAR-10, CIFAR-100 (Krizhevsky 2009), and STL-10 (Coates, Ng, and Lee 2011). |
| Dataset Splits | No | The paper mentions evaluating on CIFAR-10, CIFAR-100, and STL-10 datasets but does not explicitly state the training/validation/test splits used for these datasets within the paper's text. |
| Hardware Specification | Yes | We evaluated the total pre-training duration of 3SAT versus other competing self-AT methods on a single RTX3090 GPU. |
| Software Dependencies | No | 3SAT is built upon the BYOL (Grill et al. 2020) training script implemented by solo-learn(da Costa et al. 2022), and we strictly adhere to the settings in solo-learn for all optimizer configurations, augmentations, and projection head structures. However, specific version numbers for solo-learn or other key software dependencies are not provided. |
| Experiment Setup | Yes | We chose 256 as the batch size and performed 1000 epochs of pre-training. On the CIFAR-10 with STL-10 dataset the warm-up parameter W is set to 0, and on the CIFAR-100 dataset the warm-up parameter W is set to 200. To generate adversarial perturbations for adversarial training, we used the ℓ PGD attack (Madry et al. 2017) and followed all hyperparameters used in Dyn ACL (Luo, Wang, and Wang 2023). To speed up convergence, we only ran 5 steps of PGD in the pre-training stage. Under all evaluation methods, we only perform finetuning for 25 epochs. |