Clustered federated learning enhanced by DAG-based blockchain with adaptive tip selection algorithm

Federated learning (FL) enables machine learning on distributed data while preserving client privacy. However, FL faces challenges such as device heterogeneity, central server vulnerabilities, and non-independent and identically distributed data. To address these challenges, researchers proposed an asynchronous and decentralized clustered FL (CFL) using a directed acyclic graph (DAG)-based blockchain, called specializing DAG FL (SDAGFL). However, SDAGFL consumes high communication and storage resources, posing a substantial burden on devices with limited resources. To overcome these limitations, we propose a novel CFL framework called DAG-CFL. DAG-CFL consists of a server layer with multiple servers implementing DAG-based blockchain and a client layer. Within this framework, we propose an adaptive tip selection algorithm (ATSA) to select the most suitable tip nodes for model aggregation. The analysis indicates that DAG-CFL significantly reduces communication and storage resource consumption on the client side compared with SDAGFL. In addition, the convergence of DAG-CFL and the time and space complexity of ATSA are analyzed to show the effectiveness of DAG-CFL. We evaluate DAG-CFL and ATSA on cluster-wise MNIST and CIFAR-10 datasets. The results show that DAG-CFL achieves comparable performance to the best CFL baseline method while eliminating the need for a predefined number of clusters. Notably, DAG-CFL achieves an 8% increase in accuracy compared with SDAGFL. The experiment results also show the robustness of DAG-CLF in various data distribution shift scenarios and indicate that ATSA can effectively cluster clients with a modularity value of 0.66 for the MNIST dataset and 0.71 for the CIFAR-10 dataset.