Diabetic retinopathy and glaucoma are severe retinal disorders that can lead to irreversible vision loss if not detected and treated early. Deep learning models are commonly used to assess the severity of this condition, but they often require transmitting large amounts of sensitive data to a central server, raising privacy concerns. Federated learning addresses these concerns by keeping data decentralized, but privacy risks can still arise during the transmission of model parameters. Moreover, federated learning in heterogeneous environments may lead to longer convergence times and unstable training due to varying client capabilities. To mitigate these issues, we propose a method that combines federated learning with Paillier homomorphic encryption and a dynamic client selection mechanism. The Paillier encryption ensures data confidentiality by securely encrypting sensitive information on the client side before processing. The client selection mechanism evaluates clients based on their hardware capabilities, computational resources, and potential contribution to model updates, reducing convergence time and ensuring stable training in heterogeneous environments. Additionally, we utilize Generative Adversarial Network based augmentation to effectively address class imbalance and enhance model performance. Our approach employs lightweight models including MobileNetV2, ShuffleNet, EfficientNet-B0, and ResNet-18, making it particularly effective for diabetic retinopathy detection. Using the APTOS 2019 and Glucoma datasets from Kaggle, we demonstrate that the method achieves high accuracy, preserves privacy, and addresses challenges posed by heterogeneous hardware environments, providing a robust solution for privacy-preserving diabetic retinopathy detection.