Fast and Reliable Evaluation of Adversarial Robustness with Minimum-Margin Attack

Authors: Ruize Gao, Jiongxiao Wang, Kaiwen Zhou, Feng Liu, Binghui Xie, Gang Niu, Bo Han, James Cheng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Compared with AA, our method achieves comparable performance but only costs 3% of the computational time in extensive experiments. The reliability of our method lies in that we evaluate the quality of adversarial examples using the margin between two targets that can precisely identify the most adversarial example.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, The Chinese University of Hong Kong 2School of Mathematics and Statistics, The University of Melbourne 3RIKEN-AIP 4Department of Computer Science, Hong Kong Baptist University.
Pseudocode Yes Algorithm 1 MM Attack; Algorithm 2 Adversarial Training of MM attack.
Open Source Code Yes The code of our MM attack is available at github.com/Sjtubrian/MM-attack.
Open Datasets Yes We conducted experiments on CIFAR-10, SVHN and CIFAR-100. ... The CIFAR-10 dataset, the SVHN and the CIFAR-100 dataset can be downloaded via Pytorch.
Dataset Splits No The paper mentions using "the performance of the best checkpoint model (results at epoch 60)" which implies a validation process for model selection, but it does not explicitly provide details about a specific validation dataset split (e.g., percentages or counts) or refer to a standard validation split for the datasets used.
Hardware Specification Yes We implement all methods on Python 3.7 (Pytorch 1.7.1) with an NVIDIA Ge Force RTX 3090 GPU with AMD Ryzen Threadripper 3960X 24 Core Processor.
Software Dependencies Yes We implement all methods on Python 3.7 (Pytorch 1.7.1) with an NVIDIA Ge Force RTX 3090 GPU with AMD Ryzen Threadripper 3960X 24 Core Processor.
Experiment Setup Yes The training setup follows previous works (Madry et al., 2018; Zhang et al., 2019) that all networks are trained for 100 epochs using SGD with 0.9 momentum. The initial learning rate is 0.1 (0.01 for SVHN), and is divided by 10 at epoch 60 and 90, respectively. The weight decay is 0.0002 (0.0035 for SVHN).