Privacy-Preserving Federated Learning for Retinal Disease Diagnosis Using Paillier Homomorphic Encryption With Multiple Encryption Keys

Medical image analysis is crucial for diagnosing retinal diseases like diabetic retinopathy (DR), Retinal Vein Occlusion (RVO), and Glaucoma. This study aims to develop a secure and accurate privacy-preserving federated learning framework for retinal disease diagnosis. Federated Learning (FL) enables collaborative training without sharing raw data, mitigating privacy concerns in deep learning. However, inference attacks on model updates pose significant risks. Homomorphic encryption (HE) is widely used for privacy preservation in FL but is limited by requiring all encrypted data to use a single encryption key, increasing vulnerability. To address this, we propose a privacy-preserving FL framework integrating paillier HE with multiple encryption keys (PPFLMPHE). Each client encrypts model parameters using all client’s public keys before transmission, ensuring data security even if one key is compromised. The server aggregates encrypted local parameters per key while preserving privacy, ensuring only authorized clients can decrypt it. Additionally, we enhance classification performance and address class imbalance using data augmentation techniques such as rotation, flipping, shifting, and zooming. Our approach is evaluated on retinal image datasets, including APTOS 2019, RVO, and Glaucoma, using pre-trained models such as VGG16, ResNet50, DenseNet121, MobileNetV2, and InceptionV3. Among all models, ResNet50 provides the highest performance with 98.90% accuracy and 98.70% F1-score for DR, while DenseNet121 achieves 96.66% accuracy for RVO and InceptionV3 achieves 98.32% accuracy for Glaucoma. These results demonstrate that the proposed method effectively balances privacy preservation and model performance, offering a robust and secure FL framework for retinal disease diagnosis.