Maximum Mean Discrepancy Test is Aware of Adversarial Attacks

Authors: Ruize Gao, Feng Liu, Jingfeng Zhang, Bo Han, Tongliang Liu, Gang Niu, Masashi Sugiyama

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

Reproducibility Variable Result LLM Response
Research Type Experimental We verify detection methods on the Res Net-18 and Res Net-34 trained on the CIFAR-10 and the SVHN. We also validate performance of SAMMD on the large network Wide Res Net (WRN-32-10) (Zagoruyko & Komodakis, 2016) and the large dataset Tiny-Imagenet. Configuration of all experiments is in Appendix E. Detailed experimental results are presented in Appendix F.
Researcher Affiliation Academia 1Department of Computer Science, Hong Kong Baptist University 2Department of Computer Science and Engineering, The Chinese University of Hong Kong 3De SI Lab, AAII, University of Technology Sydney 4RIKEN-AIP 5TML Lab, University of Sydney 6Graduate School of Frontier Sciences, University of Tokyo.
Pseudocode Yes Algorithm 1 The SAMMD Test
Open Source Code Yes The code of our SAMMD test is available at github.com/Sjtubrian/SAMMD.
Open Datasets Yes We verify detection methods on the Res Net-18 and Res Net-34 trained on the CIFAR-10 and the SVHN. We also validate performance of SAMMD on the large network Wide Res Net (WRN-32-10) (Zagoruyko & Komodakis, 2016) and the large dataset Tiny-Imagenet.
Dataset Splits No The paper mentions using 'CIFAR-10 training set' and 'CIFAR-10 testing set' but does not explicitly provide specific training/validation/test split percentages, sample counts, or explicit mention of a validation set in the main text.
Hardware Specification No The paper states 'Configuration of all experiments is in Appendix E.' but does not include specific hardware details like GPU/CPU models or memory in the main text.
Software Dependencies No The paper does not specify software dependencies with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8').
Experiment Setup Yes For 6 different attacks, FGSM, BIM, PGD, AA, CW and Square (Non-IID(b)), we report the test power of all tests when SY are adversarial data (L norm ϵ = 0.0314; set size = 500) in Figure 4b.