Towards Better Robust Generalization with Shift Consistency Regularization

Authors: Shufei Zhang, Zhuang Qian, Kaizhu Huang, Qiufeng Wang, Rui Zhang, Xinping Yi

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments have been conducted to demonstrate the effectiveness of our proposed SCR method over a variety of datasets, e.g., CIFAR-10, CIFAR-100, and SVHN. It shows our method could be able to improve the robustness over the recent state-of-the-art methods substantially.
Researcher Affiliation Academia 1School of Advanced Technology, Xi an Jiaotong-Liverpool University, China. 2School of Electrical Engineering, Electronics and Computer Science, University of Liverpool, UK. 3School of Science, Xi an Jiaotong Liverpool University, China.
Pseudocode Yes Algorithm 1 Adversarial Training with SCR
Open Source Code No The paper does not provide any explicit statement or link indicating the availability of open-source code for the described methodology.
Open Datasets Yes Extensive experiments have been conducted to demonstrate the effectiveness of our proposed SCR method over a variety of datasets, e.g., CIFAR-10, CIFAR-100, and SVHN.
Dataset Splits No The paper refers to "training" and "test" datasets, for example: "The accuracy gap difference using |Acc(Sadv t , Yt) Acc(Sadv d , Yd)| |Acc(St, Yt) Acc(Sd, Yd)|, where St is the test set, Yt denotes the set of the corresponding labels of test samples in St, Sadv t is the set of adversarial examples of the test set St, and Acc(Sadv t , Yt) is the accuracy of test adversarial data set Sadv t with labels Yt.", but does not explicitly provide details about a validation split.
Hardware Specification No The paper does not specify any hardware details (e.g., CPU, GPU models, or cloud computing resources) used for the experiments.
Software Dependencies No The paper mentions software components like "SGD" and "Wide ResNet-28-10" (as a model architecture), but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes Specifically, for FS, we follow (Zhang & Wang, 2019) on CIFAR-10 and CIFAR-100. We train 200 epochs using SGD with momentum 0.9, weight decay 5 10 4, and initial learning rate 0.1. The learning rate decays at epoch 60 and 90 with the rate 0.1; for SVHN, the initial learning rate is 0.01 while the other settings remain the same as those on CIFAR-10 and CIFAR-100. For AT and TRADES, the total training epoch is 100 with the initial learning rate 0.1 for CIFAR-10 and CIFAR-100 and 0.01 for SVHN. The learning rate decays at 60 epoch with decay rate 0.1. The trade-off parameter is 0.01 on CIFAR-10 and SVHN and 0.0001 on CIFAR-100 for our proposed model. The iteration number is empirically set as 3 to compute the regularization and update step is 4/255. All other hyperparameters are the same as the baseline methods.