GAN-Augmented Attention-Based CNN for Non-IID Federated Diabetic Retinopathy Classification

Aishwarya Mane; Swati Shekapure

doi:10.15157/ijitis.2026.9.2.742-770

Authors

Aishwarya Mane Department of Computer Engineering, Marathwada Mitra Mandal’s College of Engineering, Pune and Savitribai Phule Pune University, Pune, India https://orcid.org/0009-0002-2739-4866
Swati Shekapure Department of Computer Engineering, Marathwada Mitra Mandal’s College of Engineering, Pune and Savitribai Phule Pune University, Pune, India https://orcid.org/0000-0002-9555-2275

DOI:

https://doi.org/10.15157/ijitis.2026.9.2.742-770

Keywords:

Diabetic Retinopathy, Federated Learning, Data Augmentation, Deep Learning, Generative Adversarial Network, Efficient-Net, FedProx, Non-IID Learning, Attention Mechanism

Abstract

Scalable and privacy-preserving diagnostic models are essential for diabetic retinopathy "DR" screening across multi-centre healthcare institutions, where data are heterogeneous and non-identically distributed (non-IID). Although centralised deep learning methods achieve high accuracy, they are impractical due to privacy constraints and cross-institutional data variability. To address this, we propose an attention-enhanced EfficientNet-B0 integrated with Federated Proximal Optimisation (FedProx) and GAN-based minority augmentation under simulated label-skew non-IID conditions. Experiments on APTOS 2019 and IDRiD datasets, partitioned into four federated clients, achieved 77.6% accuracy and 0.91 macro-AUC. The framework improved minority recall by 5–8% and reduced convergence variance by 44%, demonstrating stable and practical federated DR grading for real-world clinical deployment. The model was trained over 50 communication rounds across four simulated federated clients under controlled label-skew partitioning. Compared to FedAvg under identical non-IID settings (72.1% accuracy, 0.85 macro-AUC), the proposed framework demonstrates statistically significant improvement (p < 0.05).