Random Normalization Aggregation for Adversarial Defense

Authors: Minjing Dong, Xinghao Chen, Yunhe Wang, Chang Xu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on various models and datasets, and demonstrate the strong superiority of proposed algorithm.In this section, we provide sufficient evaluation of RNA module on various models and datasets.
Researcher Affiliation Collaboration Minjing Dong1, Xinghao Chen2, Yunhe Wang2, Chang Xu1 1School of Computer Science, University of Sydney 2Huawei Noah s Ark Lab mdon0736@uni.sydney.edu.au, xinghao.chen@huawei.com, yunhe.wang@huawei.com, c.xu@sydney.edu.au
Pseudocode Yes Algorithm 1 Random Normalization Aggregation with Black-box Adversarial Training
Open Source Code Yes The Py Torch code is available at https://github.com/Uni Serj/ Random-Norm-Aggregation and the Mind Spore code is available at https: //gitee.com/mindspore/models/tree/master/research/cv/RNA.
Open Datasets Yes CIFAR-10/100 We first conduct experiments on CIFAR-10/100 [31] datasets, which contain 50K training images and 10K testing images with size of 32 32 from 10/100 categories. The networks we use are Res Net-18 [31] and Wide Res Net-32 (WRN) [32].The effectiveness of proposed RNA is also evaluated on Image Net [35], which contains 1.2M training images and 50K testing images with size of 224 224 from 1000 categories.
Dataset Splits No The paper describes training and testing image counts for CIFAR-10/100 and ImageNet, but does not explicitly provide details about a validation dataset split, specific percentages for train/val/test, or mention cross-validation.
Hardware Specification Yes The experiments are performed on one V100 GPU using Pytorch [33] and Mindspore [34].The experiments are performed on eight V100 GPUs.
Software Dependencies No The paper mentions 'Pytorch [33]' and 'Mindspore [34]' as software used but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes The SGD optimizer with a momentum of 0.9 is used. The weight decay is set to 5 10 4. The initial learning rate is set to 0.1 with a piecewise decay learning rate scheduler. All the baselines are trained with 200 epochs with a batch size of 128. The PGD-10 with ϵ = 8/255 and step size of 2/255 is adopted in the adversarial training setting. The SGD optimizer with a momentum of 0.9 is used. The weight decay is set to 1 10 4. The initial learning rate is set to 0.02 with a cosine learning rate scheduler. We load a pretrained Res Net-50 and then adversarailly train the network for 60 epochs with a batch size of 512. The PGD-2 with ϵ = 4/255 is adopted in the adversarial training setting.