MultiGuard: Provably Robust Multi-label Classification against Adversarial Examples

Authors: Jinyuan Jia, Wenjie Qu, Neil Gong

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we evaluate our Multi Guard on VOC 2007, MS-COCO, and NUS-WIDE benchmark datasets. 4 Evaluation 4.1 Experimental Setup
Researcher Affiliation Academia Jinyuan Jia University of Illinois Urbana-Champaign jinyuan@illinois.edu; Wenjie Qu Huazhong University of Science and Technology wen_jie_qu@outlook.com; Neil Zhenqiang Gong Duke University neil.gong@duke.edu
Pseudocode Yes Complete algorithm: Algorithm 1 in supplementary materials shows our complete algorithm to compute the certified intersection size for an input x.
Open Source Code Yes Our code is available at: https://github.com/quwenjie/Multi Guard
Open Datasets Yes VOC 2007 [15]: Pascal Visual Object Classes Challenge (VOC 2007) dataset [15]... MS-COCO [28]: Microsoft-COCO (MS-COCO) [28] dataset... NUS-WIDE [9]: NUS-WIDE dataset [9]... We adopt the version released by [2]
Dataset Splits Yes Following previous work [43], we split the dataset into 5,011 training images and 4,952 testing images. MS-COCO [28] dataset contains 82,081 training images, 40,504 validation images, and 40,775 testing images from 80 objects. NUS-WIDE [9]... contains 154,000 training images and 66,000 testing images.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments.
Software Dependencies No The paper mentions 'Adam optimizer' and 'ASL2' (with a GitHub link) but does not specify version numbers for any software dependencies or libraries.
Experiment Setup Yes Following [2], we set training hyperameters γ+ = 0, γ = 4, and m = 0.05. We train the classifier using Adam optimizer, using learning rate 10 3 and batch size 32.