Main Contributor: Puja Saha
Training robust AI models often faces significant hurdles due to patient data privacy regulations. This project introduces novel optimization strategies for differentially private Federated Learning (FL) in medical image segmentation, enabling secure collaborative AI development across distributed clinical datasets without direct access to sensitive patient information. Our key innovation, adaptive gradient management strategies, is meticulously designed to improve the privacy-utility trade-off, ensuring high segmentation accuracy while providing robust privacy guarantees. The outcome is a highly accurate, robust, and scalable AI solution that accelerates medical research and clinical decision-making by safely overcoming traditional data sharing barriers.