OPALA: Optimized Pruning Adaptive Learning Approach for Federated Learning Scenarios

Federated learning (FL) enables decentralized model training, allowing data privacy by processing locally on client devices. However, challenges such as high communication costs, especially in heterogeneous environments, hinder the efficient scaling of FL systems. This paper introduces OPALA, an adaptive pruning optimal learning approach designed to improve personalized Federated Learning (pFL) efficiency and accuracy. OPALA combines dynamic pruning with adaptive aggregation, adjusting model complexity based on the client’s computational capabilities and data characteristics. In this sense, OPALA significantly reduces the processing time of clients to train their models based on the computational capabilities of each client, providing higher system efficiency and responsiveness to real-time applications. The results show that OPALA significantly improves training time and data transmission by up to $53 \%$ without sacrificing the model accuracy, outperforming traditional methods. This approach offers a promising solution for efficient, scalable FL in environments with non-IID data and heterogeneous client resources.