Paper Title
FEDERATED LEARNING FOR HOSPITALS WITH DIFFERENTIAL PRIVACY AND ADAPTIVE AGGREGATION

Abstract
Federated learning (FL) has emerged as a promising paradigm for training machine learning models across multiple institutions without centralising sensitive patient data. The core problem addressed in this work is threefold: (i) enabling privacy compliant collaborative training under strict HIPAA constraints;(ii) handling non-IID data distributions arising from heterogeneous patient demographics and disease prevalence across institutions; and (iii) preventing model drift caused by low-quality or unstable hospital updates. The key innovation of this work is a trust-score-driven adaptive aggregation mechanismthat weights each hospital’s contribution jointly by data quality,update consistency, and a dynamically updated trust score—going substantially beyond the data-size-only weighting used instandard FedAvg. Differential privacy (DP) based on the Opacuslibrary is integrated to provide formal (ε, δ)-DP guarantees, ensuring that individual patient records cannot be inferred fromtransmitted model updates. Experiments on the publicly available. Chest X-Ray Pneumonia dataset [3] distributed non-IID acrossfive virtual hospitals show that an EfficientNet-B0 backbone(4.34 million parameters) achieves a peak global accuracy of97.78% and AUC-ROC of 0.9773 within two communication rounds, stabilising at 95.56% across subsequent rounds. These results outperform standalone centralised CNN baselines (92.6%)and standard deep-learning approaches (94.4%), while achieving60 ms inference latency suitable for real-time clinical decisionsupport. Keywords - Federated Learning, Differential Privacy, Adaptive Aggregation, Efficientnet, Pneumonia Detection, Non-IID Data, Healthcare AI, HIPAA, Opacus, Flower Framework