Local updates and client selection strategies for federated learning systems

Federated Learning (FL) has emerged as a promising approach for enabling collab-orative machine learning across distributed clients while preserving data privacy. However, FL systems face significant challenges due to non-IID data distributions, system hetero-geneity, and communication efficiency limitations. This thesis proposes novel strategies to enhance the performance of FL through adaptive local update schemes and intelligent client selection mechanisms. First, we conduct an empirical evaluation of the impact of lo-cal update configurations on the convergence and generalization of popular FL algorithms, namely FedAVG and FedADAM. Our analysis reveals that increasing the number of local training epochs can accelerate convergence but also risks model overfitting, thereby re-ducing generalization. To mitigate this trade-off, we propose an adaptive controller that dynamically adjusts the number of local epochs using a Poisson distribution. Second, we introduce MESFLA (Model Efficiency through Selective Federated Learning Algorithm), a cluster-based client selection mechanism that considers both model weights and data size characteristics to strategically select participants for each FL training round. MES-FLA employs the CKA (Centered Kernel Alignment) algorithm to cluster clients based on model weight similarity and then selects the most relevant clients from each cluster. The adaptive local update controller dynamically balances convergence speed and model generalization, while MESFLA’s intelligent client selection optimizes performance in heterogeneous data settings. MESFLA consistently achieves higher model accuracy, faster convergence, and requires fewer communication rounds between clients and the server, demonstrating its effectiveness in non-IID data environments typical of real-world FL scenarios.