Enhancing Federated Learning Robustness with Adaptive Client Aggregation for Heterogeneous Data
Keywords:
Federated Learning, Non-IID Data, Adaptive Aggregation, Heterogeneous Data, Machine LearningAbstract
Federated learning (FL) enables collaborative machine learning across decentralised devices while preserving data privacy. However, its performance deteriorates when client data is heterogeneous and non IID (not identically distributed), leading to slower convergence and reduced accuracy. This paper proposes an adaptive client aggregation method to enhance FL’s robustness in such scenarios. Our approach dynamically adjusts the combination of client updates using techniques like clustering clients with similar data or weighting updates based on data characteristics. We evaluated this method on the Human Activity Recognition (HAR) dataset under simulated non-IID conditions. Compared to the standard Federated Averaging (FedAvg) method, our approach achieves superior accuracy and faster convergence. This research enhances the practicality of FL for real-world applications, such as healthcare and IoT, where data diversity is prevalent.
DoI: https://doi.org/10.24321/2456.1428.202534
References
McMahan B, Moore E, Ramage D, Hampson S, y Arcas BA. Communication-efficient learning of deep networksfrom decentralized data. InArtificial intelligence and statistics 2017 Apr 10 (pp. 1273-1282). PMLR.
Mishra AK, Rai S. Comparative performance assessment of eco-friendly buildings and conventional buildings of Kathmandu valley. International Journal of Current Research. 2017;9(12):62958-73.
