Synthetic Dataset derived from Global Model Trajectory
收藏资源简介：
本研究提出了一种名为SafeFL的检测框架，旨在识别联邦学习中的恶意客户端。该框架的核心是服务器收集一系列全局模型，生成一个合成数据集，用于区分恶意模型和良性模型。该合成数据集由全局模型的轨迹生成，无需反映客户端数据的实际分布，但能有效区分模型行为。SafeFL框架包括两种变体：SafeFL-ML和SafeFL-CL，分别采用不同的方法识别恶意客户端。SafeFL-ML基于恶意模型在合成数据集上损失更高的原则，通过计算损失的中位数来识别恶意客户端；SafeFL-CL则通过聚类算法将模型分组，识别损失值分布中的异常值，从而更精确地识别恶意客户端。SafeFL在五个数据集上进行了评估，包括CIFAR-10、STL-10、Tiny-ImageNet和FEMNIST等，并与其他十种最先进的联邦学习防御方法进行了比较，结果表明SafeFL在检测恶意客户端方面表现出色。
This study proposes a detection framework named SafeFL, which aims to identify malicious clients in federated learning. The core of this framework is that the server collects a series of global models to generate a synthetic dataset for distinguishing malicious models from benign ones. This synthetic dataset is generated from the trajectories of global models, which does not need to reflect the actual distribution of client data, but can effectively distinguish model behaviors. The SafeFL framework includes two variants: SafeFL-ML and SafeFL-CL, which adopt different methods to identify malicious clients. Based on the principle that malicious models incur higher losses on the synthetic dataset, SafeFL-ML identifies malicious clients by calculating the median of loss values. SafeFL-CL, on the other hand, groups models via clustering algorithms and identifies outliers in the loss value distribution, thereby identifying malicious clients more accurately. SafeFL was evaluated on five datasets including CIFAR-10, STL-10, Tiny-ImageNet, FEMNIST and others, and compared with ten state-of-the-art federated learning defense methods. The results show that SafeFL performs excellently in detecting malicious clients.
