FedLBW Improves Federated Learning Accuracy on Non-IID Wireless Data

Federated Learning (FL) enables collaborative machine learning across distributed clients, crucial for privacy-preserving applications. However, achieving efficient model convergence in FL is particularly difficult in wireless networks due to non-independent and identically distributed (non-IID) data and frequent client dropouts. Traditional FL algorithms, like FedAvg, typically weight client updates solely by dataset size, which can introduce biases and sensitivity to outliers.To overcome these limitations, the authors propose Federated Learning with Loss-Based Weighting (FedLBW). This new aggregation method assigns weights to each client's update proportional to the inverse of its validation loss, calculated using a small proxy dataset on the server. This strategy ensures that models with lower validation loss, indicating higher reliability, exert greater influence during the aggregation process.Extensive experiments across various datasets and models (FashionMNIST, CIFAR-10, CIFAR-100) demonstrate that FedLBW achieves higher accuracy and faster convergence compared to several baseline FL algorithms, with accuracy improvements of up to 7.6% in extreme non-IID scenarios. Furthermore, FedLBW exhibits exceptional resilience to increasing dropout probabilities, maintaining significantly higher accuracy under challenging conditions. These results position FedLBW as a robust and effective solution for FL in wireless network environments.