MIBench
收藏资源简介:
MIBench是由哈尔滨工业大学(深圳)和清华大学共同创建的首个模型反转攻击和防御的综合基准。该数据集包含四个广泛认可的人脸数据集,包括Flickr-Faces-HQ (FFHQ)、MetFaces、FaceScrub和CelebFaces Attributes (CelebA)。这些数据集主要用于预训练辅助先验和攻击目标私有数据集。数据集创建过程中,提供了低分辨率和高分辨率的两个版本,并配备了多种图像处理工具,以简化研究者的数据预处理任务。MIBench旨在解决模型反转攻击中的隐私泄露问题,通过提供一个统一、可扩展和可复现的工具箱,促进该领域的进一步研究和创新。
MIBench is the first comprehensive benchmark for model inversion attacks and defenses, co-developed by Harbin Institute of Technology (Shenzhen) and Tsinghua University. This benchmark incorporates four widely adopted facial datasets, including Flickr-Faces-HQ (FFHQ), MetFaces, FaceScrub, and CelebFaces Attributes (CelebA). These datasets are primarily used for pre-training auxiliary priors and as target private datasets for attacks. During the development of MIBench, two variants (low-resolution and high-resolution) are provided, along with a suite of image processing tools to streamline researchers' data preprocessing workflows. MIBench aims to address privacy leakage issues in model inversion attacks, and facilitate further research and innovation in this domain by offering a unified, scalable, and reproducible toolkit.




