Speaker: 孔令臣教授，北京交通大学数学与统计学院           

Inviter: 中国科学院数学院学术报告会组委

Title: A Novel Conformal Prediction Framework for Differential Privacy

Language: Chinese

Time & Venue: 2025.01.03 10:40-11:50  数学院南楼N204（数学院交流报告会）

Abstract: As privacy and reliability become increasingly critical in modern statistical machine learning, we introduce a novel differential private conformal prediction (DPCP). It is a framework for constructing model-free private prediction sets to safely enable uncertainty quantification. Compared to existing methods, it has lower computational cost and more efficient data utilization, while maintaining rigorous theoretical guarantees on both privacy and coverage. Moreover, it produces more precise and less conservative sets of predictions. We further analyze the efficiency of DPCP within the frameworks of empirical risk minimization, demonstrating their robustness and adaptability. Numerical experiments on real-world datasets validate the practical effectiveness of our approach.