Model Heterogeneous Federated Learning with Noisy Data

Each client designs its own model independently, making the task even more difficult. Existing algorithms are not efficient in solving the problem of varying noise in local clients, which is caused by the difficulty of data labeling and hitchhiking clients. In this paper, we address the challenging problem of federated learning with noisy and heterogeneous clients. We propose a new solution, Federated Classifier Jointing (FedClassJoint), which simultaneously handles label noise and performs federated learning in a single framework. The deep neural network used for the supervised learning task consists of a feature extractor layer and a classifier layer. Additionally, we apply local feature representation learning to stabilize the decision boundary and improve the local feature extraction capability of the client. FedClassJoint is characterized by three aspects: (1) efficient communication between heterogeneous models, achieved by requiring the client to communicate with only a few fully connected layers; (2) reduction of negative impact caused by internal label noise through the use of contrastive regularization loss function (CCTR). (3) To address noisy feedback from other participants, we have designed a new client confidence reweighting scheme. This scheme adaptively assigns appropriate weights to each client classifier during the collaborative learning phase. The classifier weights are then aggregated into a decision boundary protocol on the feature space, resulting in a powerful global classifier. Our approach has been extensively tested and has proven effective in minimizing the negative impact of various noise rates and types in both homogeneous and heterogeneous federated learning settings. It consistently outperforms existing methods.