Sparse DNNs with Improved Adversarial Robustness

Authors: Yiwen Guo, Chao Zhang, Changshui Zhang, Yurong Chen

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our analyses reveal, both theoretically and empirically, that nonlinear DNN-based classifiers behave differently under l2 attacks from some linear ones. Towards shedding light on such relationships, especially for DNNs, we provide comprehensive analyses in this paper from both the theoretical and empirical perspectives. In this section, we conduct experiments to testify our theoretical results.
Researcher Affiliation Collaboration Yiwen Guo1, 2 Chao Zhang3 Changshui Zhang2 Yurong Chen1 1 Intel Labs China 2 Institute for Artificial Intelligence, Tsinghua University (THUAI), State Key Lab of Intelligent Technologies and Systems, Beijing National Research Center for Information Science and Technology (BNRis), Department of Automation, Tsinghua University 3 Academy for Advanced Interdisciplinary Studies, Center for Data Science, Peking University {yiwen.guo, yurong.chen}@intel.com pkuzc@pku.edu.cn zcs@mail.tsinghua.edu.cn
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide information about open-source code for the methodology described.
Open Datasets Yes We choose the well-established MNIST dataset as a benchmark, which consists of 70,000 28 28 images of handwritten digits. Experiments are also conducted on CIFAR-10, in which deeper nonlinear networks can be involved.
Dataset Splits No According to the official test protocol, 10,000 of them should be used for performance evaluation and the remaining 60,000 for training.
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No We also follow the training policy suggested by Caffe [17] and train network models for 50,000 iterations with a batch size of 64... (Caffe is mentioned, but no version number is given, nor are other software dependencies with versions).
Experiment Setup Yes We also follow the training policy suggested by Caffe [17] and train network models for 50,000 iterations with a batch size of 64 such that the training cross-entropy loss does not decrease any longer. Here we set m = 16, ρ = 1/3 so the achieved final percentage of zero weights should be 99.74% 1 (1 ρ)m.