AdaFedProx: A Heterogeneity-Aware Federated Deep Reinforcement Learning for Medical Image Classification

In the realm of smart healthcare, vast amounts of valuable patient data are generated worldwide. However, healthcare providers face challenges in data sharing due to privacy concerns. Federated learning (FL) offers a privacy-preserving solution by enabling collaborative model training without direct access to patient data. This decentralized approach utilizes data from diverse sources, resulting in a globally learned model with satisfactory performance across individual sites. However, federated training encounters challenges such as “system heterogeneity” in client specifications and “data heterogeneity” in imbalanced data distribution. Standard FL methods become unstable and require extensive hyperparameter tuning for optimal performance when facing heterogeneous clients, posing challenges in real-world applications. This research introduces a novel heterogeneity-aware federated learning approach named AdaFedProx to address the performance degradation resulting from heterogeneity. AdaFedProx utilizes a reinforcement learning-driven strategy, introducing a tailored proximal term into the objective function to incorporate partial client information in heterogeneous settings. Dynamically determining the proximal term based on individual client states, including critical characteristics like data distribution, system specification, and performance feedback, empowers AdaFedProx to make regularized decisions throughout the federated training. AdaFedProx consistently outperforms four state-of-the-art methods on three medical datasets across diverse heterogeneity settings (0%, 50%, and 75% straggler). The code can be found at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Pranabiitp/AdaFedProx</uri>.