CIFAR10
收藏资源简介:
本研究使用CIFAR10数据集,这是一个广泛用于图像分类任务的标准数据集,包含10个类别的60000张32x32彩色图像。研究团队通过预训练的神经网络记录批量归一化(BN)层的统计数据,进而生成与原始数据集统计分布相匹配的合成数据集。该合成数据集旨在用于隐私保护的机器学习应用,通过调整优化步骤的数量,可以在保持一定隐私的同时,实现合理的分类性能。
This study utilizes the CIFAR-10 dataset, a standard benchmark widely adopted for image classification tasks, which contains 60,000 32×32 color images across 10 classes. The research team recorded the statistical parameters of batch normalization (BN) layers from pre-trained neural networks, thus generating a synthetic dataset that matches the statistical distribution of the original dataset. This synthetic dataset is intended for privacy-preserving machine learning applications; by tuning the number of optimization steps, reasonable classification performance can be achieved while maintaining a certain level of privacy.




