Phase Transition from Clean Training to Adversarial Training

Authors: Yue Xing, Qifan Song, Guang Cheng

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate this conjecture in linear regression models, and conduct comprehensive experiments in deep neural networks.
Researcher Affiliation Academia Yue Xing Department of Statistics Purdue University xing49@purdue.edu Qifan Song Department of Statistics Purdue University qfsong@purdue.edu Guang Cheng Department of Statistics University of California, Los Angeles guangcheng@ucla.edu
Pseudocode No The paper refers to an 'algorithm' and describes steps for approximation in Section 6 and Appendix A, but it does not present a formal pseudocode or algorithm block.
Open Source Code No We provide the github links for the original code we modified from.
Open Datasets Yes For datasets CIFAR-10, CIFAR-100, SVHN, we modified the code of Rice et al. (2020) for our implementation. ... For MNIST, we implement a CNN with two convolution layers and two fully connected layers.
Dataset Splits No We use an SGD optimizer with batch size 128 to train on the full training set for 200 epochs. ... After training 200 epochs, we find the epoch with the smallest adversarial testing loss as the final model for early stopping.
Hardware Specification No The paper mentions training times (e.g., 'it takes 20 minutes to train the CIFAR-10 dataset for clean training but takes 10 hours to complete adversarial training') but does not specify the type of hardware used, such as GPU or CPU models.
Software Dependencies No The paper states 'we modified the code of Rice et al. (2020) for our implementation' but does not list specific software dependencies with version numbers.
Experiment Setup Yes If there is no specification, we use an SGD optimizer with batch size 128 to train on the full training set for 200 epochs. The learning rate is initialized as 0.1 for CIFAR and 0.01 for SVHN, and is divided by 10 at the 100th and 150th epochs. We consider Pre Act Res Net18 and Wide Res Net34 as those in Rice et al. (2020).