DFedCL: Decentralized Federated Collaborative Learning with Privacy Protection

Federated Learning (FL) is a new machine learning paradigm that enables multiple clients to collaboratively train a global model without sharing raw data. To achieve privacy protection, existing FL methods incorporate artificial noise into client model updates to ensure differential privacy (DP). However, these DP-FL approaches are vulnerable to single-point failures at the server node and suffer significant performance declines in non-independent and identically distributed (non-IID) scenarios. To mitigate the research gap, we propose the Decentralized Federated Collaborative Learning (DFedCL.) DFedCL framework is designed as a fully decentralized system where clients can directly share updated models without a server node. Within this framework, each client maintains a private model locally and uses a sharing model for information exchange. These two models are collaboratively trained and updated in each learning iteration based on the noise-added gradients generated. To improve the model generalization ability and enhance model learning performance in non-IID scenarios, we further propose the SAM-based update correction, applied by the sharing model exchange. We evaluate our proposed DFedCL through MNIST and CIFAR-10 datasets in different non-IID degrees, the experimental results show the DFedCL achieves state-of-the-art performance.