Enhancing global model accuracy in federated learning with deep neuro-fuzzy clustering cyclic algorithm

Abstract

In recent years, with the increasing importance of privacy protection, many laws and regulations have standardized data usage, requiring companies to obtain user consent to access personal data. This has become more challenging for models that require large amounts of data for training. Therefore, the concept of federated learning was proposed in 2016, aiming to train models with different clients without sharing data to ensure data privacy. However, federated learning faces several challenges, including heterogeneous devices, data security, data heterogeneity, communication costs, and training time costs. This study focuses on addressing the issue of data heterogeneity, where the data distribution among participating clients differs significantly, leading to poor performance of the aggregated model after training. To tackle this problem, we propose a federated clustering cyclic algorithm, which involves two-step clustering of clients to make the data distribution of each cluster approach independent and identically distributed. We also introduce deep neural fuzzy methods to handle fuzzy, uncertain, or incomplete data. According to experimental results, the proposed deep neuro-fuzzy clustered cyclic algorithm outperforms methods such as FedAvg, FedProx, and CyclicFL on various non-IID datasets, with accuracy approaching that of centralized learning in certain experiments. This indicates that the deep neural fuzzy methods and clustering cyclic algorithm DNCC presented in this study can improve the accuracy of global models, especially in increasingly non-IID scenarios. Furthermore, we extend this method to big data processing to cope with more complex data environments.

Keywords

• Federated learning
• Non-IID data
• Deep neuro-fuzzy clustering
• Cyclic training
• Model convergence