A Review of Privacy-Preserving Deep Learning for Automated Diabetic Retinopathy Detection Using Hybrid CNN–Transformer Models

Abstract

Diabetic Retinopathy (DR) is a serious retinal disease that occurs due to diabetes and is one of the major causes of vision loss globally. While recent advances in computer vision and deep learning have significantly enhanced automated retinal image analysis for early DR detection, centralized learning methods introduce important considerations of patient privacy, regulatory standards, data leakage and cyber-security risks. In this review, we explore the privacy-preserving deep learning frameworks like Federated Learning (FL), Differential Privacy (DP), Homomorphic Encryption (HE), and Secure Multi-Party Computation (SMPC) for automated diabetic retinopathy (DR) detection using CNN and hybrid CNN-Transformer architectures. This study investigates the present status of decentralized medical imaging systems for collaborative model learning without sharing sensitive patient data and critically analyzes the evolution of the systems. Further, the research delves into complex hybrid models, which integrate Convolutional Neural Networks (CNNs) with Vision Transformers (ViTs), to achieve both local and global context understanding for retinal disease classification. A comparative study of recent state-of-the-art models shows that privacy-preserving federated models achieve over 95% diagnostic accuracy, enjoy formal privacy guarantees, and are resistant to gradient inversion and membership inference attacks and poisoning attacks. Furthermore, several different aspects of communication efficiency, heterogeneity-aware federated optimization, embedding explainable artificial intelligence (XAI) for clinical deployment, and problems faced in multi-institutional healthcare environments are discussed. Finally, the future research directions of large-scale federated learning systems, trustworthy explainable AI, fairness-aware learning, secure medical security using blockchain, and lightweight edgedeployable Transformer models are pointed out for future intelligent ophthalmic diagnostic systems. Keywords: Diabetic Retinopathy Detection, Federated Learning, Privacy-Preserving Deep Learning, Hybrid CNN– Transformer Networks, Differential Privacy, Homomorphic Encryption, Vision Transformers in Healthcare, Explainable Medical AI