Optimal Model Transfer and Dynamic Parameter Server Selection for Efficient Federated Learning in IoT-Edge Systems With Non-IID Data

Distributed and privacy-preserving federated learning (FL) has been associated with edge computing systems for developing intelligent IoT applications. However, collecting data individually in each FL node may result in non-independent and identically distributed (non-IID) training data, which can significantly impair FL performance. To address this, we propose a model transfer approach that allows a FL client to train its model on more datasets, collectively forming an IID virtual dataset. In the proposed approach, FL clients are classified as transfer-demand or aggregation-inclined based on their local model’s experienced data label distributions. Transfer-demand clients transfer their models to helper clients that can support their lacking labels, while aggregation-inclined clients have enough data labels in their training models, and thus, they participate in the model aggregation. However, two or more transfer-demand clients may contend with the same helper client. To resolve the contention, we apply the minimum-cost maximum-matching (MCMM) framework and integer linear programming to find the optimal solution. To minimize model transmission costs among FL clients, we use a Steiner tree-based solution to dynamically allocate a parameter server that aggregates the local models from clients. Finally, we perform extensive simulation experiments with different problems, and our proposed approach significantly outperforms baseline methods, in the case of MNIST achieves 99.02% accuracy and reduces the communication cost within the range of 20% to 66% and in the case of CIFAR-10 dataset, enhances the accuracy at least 24% and communication cost in the range of 29.48% to 63.09% in comparison with non-IID baseline approaches.