Research on Privacy, Fairness and Utility Improvement of Deep Learning Models
Abstract
Deep learning (DL) has emerged as a transformative paradigm in healthcare care, driving advances in medical image analysis, automated diagnosis, and clinical decision support. Despite these successes, current DL models face critical limitations in reliability, privacy, fairness, and utility, particularly when deployed in sensitive applications. This dissertation systematically investigates these challenges and proposes novel methods to improve reliability, privacy, fairness, and accuracy. To address the need for clinically reliable detection of diabetic retinopathy, we first develop a transfer learning and ensemble-based diabetic retinopathy (DR) classification system evaluated using the quadratic weighted kappa (QWK) metric. Using the EfficientNet-B3 model and two-stage ensemble learning, the proposed method achieves QWK scores of 0.901, 0.967, and 0.944 on the EyePACS, APTOS, and Messidor-2 datasets, respectively—surpassing existing state-of-the-art approaches. The framework demonstrates robustness across imaging conditions and datasets, providing an efficient and scalable pipeline for the screening of DR. To enhance data privacy while preserving model accuracy and fairness, we introduce DP-SGD-Global-Adapt-V2-S, a differentially private stochastic gradient descent algorithm that integrates a step-decay noise multiplier and an adaptive gradient clipping threshold. Across the MNIST, CIFAR-10, and CIFAR-100 benchmarks, the proposed method achieves accuracy improvements of up to 8.4% and fairness gains of 6.7% compared to standard DP-SGD while maintaining an equivalent privacy budget (ϵ = 1). Further analysis demonstrates that step-decay noise scheduling stabilizes convergence and improves generalization under privacy constraints, achieving an optimal trade-off between privacy, fairness, and utility. Finally, to improve performance balance across multi-modal data and overall accuracy of the multi-modal model, we design a Balanced Soft Mixture-of-Experts (SMoE) framework for glaucoma detection that fuses fundus images and OCT scans. The SMoE model dynamically reweights modality contributions using a gating network to mitigate the modality imbalance. Across the FairDomain, FairVision, and HarvardGF datasets, the proposed system achieves AUC scores exceeding 0.96, outperforming uni-modal and traditional fusion methods by 4–7%. Together, the results demonstrate that it is possible to design DL systems that are reliable, accurate, privacy-preserving, and fair; This dissertation provides theoretical and practical advances, establishing a foundation for the reliable and ethical deployment of deep learning technologies in real-world applications such as healthcare and 3d printing.