Adaptive Federated Learning with Generative AI for Privacy-Preserving Healthcare
Adaptive Federated Learning with Generative AI for Privacy-Preserving Healthcare

Summary/Abstract: Advancements in healthcare monitoring have led to significant data generation, raising urgent concerns about privacy and data sharing. Federated Learning (FL) has emerged as a privacy-preserving solution by enabling decentralized model training across institutions. This study presents an Adaptive Federated Learning framework across four simulated hospital clients with heterogeneous healthcare data. We evaluate three FL strategies - FedAvg, FedProx, and an adaptive FedOpt - highlighting the efficacy of dynamically tuned aggregation using the Adam optimizer. To improve model generalization while preserving privacy, we integrate synthetic datasets generated via Generative AI (ChatGPT and Mostly AI), simulating realistic yet anonymized clinical scenarios. The system includes continuous performance monitoring, particularly addressing the underperformance of one client (Client A4 and B4), ensuring fairness through personalized tuning. Our approach achieves up to 91.3% accuracy and 0.88 F1-score, demonstrating robust performance in non-IID data settings. This work uniquely combines Generative AI with adaptive FL, contributing a scalable, privacy-conscious learning paradigm for future healthcare AI deployments.

• Details
• Contents