Local-Global Codistillation Framework Federated Learning (LG-FedDistill)

Federated Learning is a machine learning framework that allows multiple agents to train models on localized datasets collectively without explicitly sharing their raw data, ensuring privacy over data. However, in real deployment scenarios, client datasets are often imbalanced and exhibit significant statistical heterogeneity (non-IID conditions), which complicates the training process. Locally biased updates can affect global aggregation, reducing training stability and limiting the reliability and generalization capability of the final global model. Conventional aggregation methods, such as Federated Averaging (FedAvg), assign equal importance to all client updates, making them vulnerable to biases caused by skewed local data distributions. In this work, we introduce a Local-Global Co-Distillation Federated Learning framework (LG-FedDistill) designed to enhance aggregation robustness across a broad range of data skew conditions. The proposed framework arranges the exchange of prediction logits among participants and employs a composite objective function that jointly leverages local supervisory signals and distilled global knowledge, thereby promoting more stable and reliable convergence in heterogeneous federated environments. The framework is evaluated on the COVID-19 Radiography Database under multiple skew levels.