基于生成式数据增强的联邦多标签弱监督图神经网络及其应用研究

Generative Data Augmentation Based Federated Multi-Label Weakly Supervised Graph Neural Networks and Their Applications

摘要: 隐私敏感领域的图像分类问题面临诸多挑战，尤其设备磨损等因素会在多标签电力场景分类任务中引入标签噪声，而区域差异会导致不同区域的本地数据集上出现类别不平衡问题。尽管联邦学习(Federated Learning, FL)是一种有效的隐私保护解决方案，但很少有联邦学习框架专门针对多标签电力场景分类中的标签噪声和类别不平衡问题来研究有效的解决方案。在此背景下，本文提出了一种基于生成式数据增强的联邦多标签弱监督图神经网络(Federated Multi-label Graph Neural Network with Generative Data Augmentation for Noisy Electricity Image Classification, FedMGNN)，用于监督的标签信息包含噪声的电力图像分类。该网络在本地和全局模型构建过程中融入两方面的关键创新，旨在从本地模型和全局模型两方面协同解决标签噪声和类别不平衡问题。首先，设计了一种客户端的本地识别模型，将基于生成模型的数据增强算法与多标签对比学习框架相融合，从而同时应对两个关键挑战：(1) 通过合成代表性样本弥补少数类别的数据不足问题，从数据上缓解类别不平衡；(2) 利用本地模型和接收的全局模型构建对比学习框架，通过对比损失和样本筛选策略来过滤含噪声标签的样本，减轻标签噪声的影响。其次，设计了一种全局聚合策略，采用类别不平衡感知加权方法，将分布式本地模型整合为统一的多标签分类框架。最后，通过在不同标签噪声比例的联邦场景图像数据集上的实验验证了所提出的FedMGNN模型的有效性。

Abstract: The image classification problems in privacy-sensitive domains faces numerous challenges, particularly as factors like equipment wear and tear introduce label noise in multi-label electricity scene classification tasks, and regional differences lead to class imbalance in local datasets from different areas. While Federated Learning (FL) is an effective privacy-preserving solution, few FL frameworks specifically address the issues of label noise and class imbalance in multi-label electricity scene classification tasks. In this context, this paper proposes a Federated Multi-label Graph Neural Network with Generative Data Augmentation for Noisy Electricity Image Classification (FedMGNN), which is designed to handle noisy label information in supervised electricity image classification. The network integrates two key innovations in both local and global model construction to collaboratively address label noise and class imbalance from both the local and global model perspectives. First, we design a local recognition model for the client, combining data augmentation based on generative models with a multi-label contrastive learning framework. This approach addresses two key challenges simultaneously: (1) by synthesizing representative samples, it alleviates the data scarcity problem of minority classes, thus mitigating class imbalance; (2) by leveraging both the local model and the received global model to construct the contrastive learning framework, it filters noisy label samples through contrastive loss and sample selection criteria, thereby reducing the impact of label noise. Secondly, we propose a global aggregation strategy that uses a class imbalance-aware weighting method to integrate distributed local models into a unified multi-label classification framework. Finally, experiments on a federated scene image dataset with varying label noise ratios validate the effectiveness of the proposed FedMGNN model.