Model Compression with Adversarial Robustness: A Unified Optimization Framework

Authors: Shupeng Gui, Haotao Wang, Haichuan Yang, Chen Yu, Zhangyang Wang, Ji Liu

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate that ATMC achieves remarkably favorable trade-offs between robustness and model compactness, we carefully design experiments on a variety of popular datasets and models as summarized in Section 3.1.
Researcher Affiliation Collaboration Department of Computer Science, University of Rochester Department of Computer Science and Engineering, Texas A&M University Ytech Seattle AI lab, Fe DA lab, AI platform, Kwai Inc
Pseudocode Yes Algorithm 1 Zero Kmeans B( U) and Algorithm 2 ATMC
Open Source Code Yes The codes are publicly available at: https://github.com/shupenggui/ATMC.
Open Datasets Yes Le Net on the MNIST dataset [59]; Res Net34 [60] on CIFAR-10 [61] and CIFAR-100 [61]; and Wide Res Net [62] on SVHN [63].
Dataset Splits No The paper mentions training and testing sets, but does not explicitly provide details about a distinct validation dataset split or its size/percentage.
Hardware Specification No The paper does not specify the hardware used for running the experiments, such as GPU models, CPU types, or memory details.
Software Dependencies No The paper mentions using PyTorch implicitly through common deep learning practices but does not specify exact versions for software dependencies like PyTorch, Python, or CUDA.
Experiment Setup Yes Unless otherwise specified, we set the perturbation magnitude to be 76 for MNIST and 4 for the other three datasets. (The color scale of each channel is between 0 and 255.) Following the settings in [32], we set PGD attack iteration numbers n to be 16 for MNIST and 7 for the other three datasets. We follow [30] to set PGD attack step size α to be min( + 4, 1.25 )/n. We train ATMC for 50, 150, 150, 80 epochs on MNIST, CIFAR10, CIFAR100 and SVHN respectively.