差分隱私保證的求和問題的實例最優裁剪
講座時間：2025-07-11 09:30講座地點：清水河校區主樓B1-104
特邀專家： Wei Dong，恒久从事差分隐私、数据宁静与数据库理论研究，在SIGMOD、PODS、CCS、IEEE S&P、VLDB、NeurIPS等國際頂級會議發表多篇論文，结果多次獲得最佳論文獎和SIGMOD博士論文獎Runner-up等重要榮譽。其研究涵蓋差分隱私機制、聯邦學習中的隱私保護、私有SQL處理等領域，並提出多種實例最優（Instance-optimal）數據發布技術。
講座內容： Many differentially private mechanisms focus on worst-case error bounds, but instance-specific guarantees—where error depends on actual data values—can provide more meaningful accuracy. In the sum estimation problem, existing methods in the shuffle model often require multiple communication rounds to achieve instance-optimal accuracy. We propose a new approach that performs both clipping and sum estimation in a single round, significantly improving efficiency while maintaining strong accuracy guarantees. Our method also extends to high-dimensional sum estimation and sparse vector aggregation, making it a practical solution for privacy-preserving data analysis.
