Efficient Privacy-Utility Optimization for Differentially Private Deep Learning Application to Medical Diagnosis

The optimization of differentially private deep learning models in medical data analysis using efficient hyper-parameter tuning is still a challenging task. In this context, we address the fundamental issue of balancing privacy guarantees with model utility by simultaneously optimizing model parameters and privacy parameters across two primary medical datasets, with additional validation on PathMNIST. Our framework encompasses both tabular data (Wisconsin Breast Cancer dataset) and medical imaging (BreastMNIST and PathMNIST), implementing four distinct optimization approaches: Grid Search, Random Search, Bayesian Optimization, and Bat Algorithm. Through extensive experimentation, we demonstrate a promising performance: achieving 93.62% accuracy with strong privacy guarantees (ε = 0.5) for tabular data, and 74.91% accuracy for medical imaging, with the Bat Algorithm discovering an unprecedented privacy level (ε = 0.293). Further validation on PathMNIST histopathology images demonstrated the framework's scalability, achieving 44.71% accuracy with privacy guarantees (ε = 2.603). Our comparative analysis reveals that different medical data types require distinct optimization strategies, with Bayesian Optimization excelling in tabular data applications and Random Search providing efficient solutions for image processing. The experiments with PathMNIST histopathology images provided valuable insights into the framework's behavior with complex medical data, revealing configuration-dependent performance variations and computational trade-offs. Our framework incorporates Pareto analysis and visualization techniques to enable systematic exploration of privacy-utility trade-offs, while early stopping mechanisms optimize privacy budget utilization. This comprehensive approach, validated across diverse medical imaging complexities and data modalities, establishes practical guidelines for implementing privacy-preserving machine learning in healthcare settings while highlighting the importance of balanced optimization strategies and computational efficiency in secure and efficient medical data analysis.