ImageNet-100 对抗样本数据

在ImageNet-100图像分类数据集上，运用PGD、FGSM对抗样本攻击算法生成，目标模型为ResNet50。随机选择与标签不一致的类别作为扰动目标。数据目录下有2000条对抗样本数据，每条数据分别对应原始图像和对抗样本图像，总共2000*2=4000条数据。每条数据使用index标记了序号，使用[clean,adv]标记数据类型，所有数据均为jpeg文件类型。在此基础上，本项目自研并验证了8种针对分类器威胁的检测和防御方法，着重研究面向常见对抗样本的生成方法带来的安全威胁。Generated via PGD and FGSM adversarial sample attack algorithms on the ImageNet-100 image classification dataset, with ResNet50 as the target model. Classes inconsistent with the original labels were randomly selected as perturbation targets. The data directory contains 2000 pairs of adversarial sample data, where each pair consists of a clean original image and its corresponding adversarial sample image, totaling 4000 data items (2000×2=4000). Each data item is assigned an index for serial numbering, and the data type is marked with [clean, adv]. All data are stored in JPEG file format. Based on this dataset, this project independently developed and validated 8 detection and defense methods against classifier threats, with a focus on the security threats induced by common adversarial sample generation techniques.

搜集汇总

背景与挑战

背景概述

该数据集基于ImageNet-100图像分类数据集生成，使用PGD和FGSM对抗样本攻击算法，以ResNet50为目标模型，包含2000条对抗样本数据及其对应的原始图像，总计4000条数据。数据集旨在研究分类器的安全威胁，并验证了相关的检测和防御方法。