Learning Adversarially Robust Sparse Networks via Weight Reparameterization
Authors: Chenhao Li, Qiang Qiu, Zhibin Zhang, Jiafeng Guo, Xueqi Cheng
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on diverse datasets demonstrate that our method achieves state-of-theart results, outperforming the current sparse robust training method and robustness-aware pruning method. |
| Researcher Affiliation | Academia | Chenhao Li1, 2, Qiang Qiu1, Zhibin Zhang1, Jiafeng Guo1, Xueqi Cheng1 1 CAS Key Lab of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China {lichenhao19b, qiuqiang, zhangzhibin, guojiafeng, cxq}@ict.ac.cn |
| Pseudocode | Yes | Algorithm 1: Pipeline of adversarial network pruning |
| Open Source Code | Yes | Our code is available at https://github.com/UCAS-LCH/Twin-Rep. |
| Open Datasets | Yes | In our main evaluations, we evaluate the robustness of VGG-16 (Simonyan and Zisserman 2015) and Wide-Res Net-28-4 (Zagoruyko and Komodakis 2016) networks on CIFAR-10 (Krizhevsky et al. 2009) and SVHN (Netzer et al. 2011) image classification datasets as most of the baseline works report their results under this setting. |
| Dataset Splits | No | The paper uses datasets like CIFAR-10 and SVHN but does not explicitly provide the training, validation, and test dataset splits needed for reproduction. It does not specify percentages or sample counts for each split. |
| Hardware Specification | No | The paper does not explicitly provide specific hardware details such as GPU or CPU models used for running experiments. |
| Software Dependencies | No | The paper mentions optimization methods like 'SGD with momentum 0.9' and specific loss functions like 'TRADES', but it does not specify any software names with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8'). |
| Experiment Setup | Yes | Optimization for all models is performed using SGD with momentum 0.9 and an initial learning rate of 0.1, which is divided by 10 at the 70-th and 85-th epoch. All models are trained for 100 epochs with a batch size of 128, and we perform pruning at the 30-th epoch. Network weights (W1 and W2) are initialized via Kaiming initialization (He et al. 2015). Besides, the weight decay factor is set to 2e-4. |