DataElixir
收藏资源简介:
DataElixir是一种新颖的数据集净化方法,旨在通过扩散模型消除中毒样本中的触发特征并恢复良性特征,从而将中毒样本转化为良性样本。该方法通过多次迭代正向和反向过程,提取原始数据集中每个样本的中间图像及其预测标签,识别标签转换中的异常样本,通过量化分布差异检测目标标签,并根据像素和特征距离选择净化图像,通过训练良性模型确定其真实标签。实验证明,DataElixir能有效缓解多种复杂攻击,同时对良性准确率的影响最小。
DataElixir is a novel dataset purification method designed to eliminate trigger features in poisoned samples and restore benign features through diffusion models, thereby transforming poisoned samples into benign ones. This method iteratively performs forward and backward processes to extract intermediate images and their predicted labels from each sample in the original dataset, identifies anomalous samples during label transitions, detects target labels by quantifying distribution differences, selects purified images based on pixel and feature distances, and determines their true labels by training a benign model. Experiments demonstrate that DataElixir effectively mitigates various complex attacks while minimally impacting benign accuracy.
数据集概述
数据集名称
DataElixir
数据集目的
用于净化受污染的数据集,以减轻基于中毒的后门攻击。
数据集方法
利用扩散模型消除触发特征并恢复良性特征,通过多次迭代的前向和反向过程,提取中间图像及其预测标签,识别异常样本,并通过量化分布差异检测目标标签,选择净化图像,并通过训练良性模型确定其真实标签。
数据集效果
在9种流行攻击上进行的实验表明,DataElixir能有效减轻各种复杂攻击,同时对良性准确度的影响最小,超越了基线防御方法的性能。
数据集引用信息
@misc{zhou2023dataelixir, title={DataElixir: Purifying Poisoned Dataset to Mitigate Backdoor Attacks via Diffusion Models}, author={Jiachen Zhou and Peizhuo Lv and Yibing Lan and Guozhu Meng and Kai Chen and Hualong Ma}, year={2023}, eprint={2312.11057}, archivePrefix={arXiv}, primaryClass={cs.CR} }




