人工智能学院格物致智论坛-Instance Optimality in Differential Privacy

讲座摘要：

￮Because of its mathematical rigor and applicability to a broad class of problems, differential privacy (DP) has now become the de facto standard for protecting personal information, widely adopted by both governments and industry.  Most classical DP mechanisms are based on the sensitivity of the target function, but many basic functions, such as sum, mean, median, max/min, as well as most SQL queries, do not have a bounded sensitivity.  For such functions, existing algorithms often take an ad hoc approach when measuring their utility.  In this talk, I will present a unified and principled framework to define the optimality of DP mechanisms for such functions, based on a slightly relaxed notion of instance optimality, and show how this can be achieved for the aforementioned problems.

主讲人简介：

易珂是香港科技大学计算机科学与工程学系教授，美国计算机学会（ACM）会士。先后获得清华大学计算机科学学士学位和杜克大学计算机科学博士学位。他的研究方向包括数据库理论与系统、数据安全与隐私、算法、数据结构以及理论计算机科学。曾两次获得 SIGMOD 最佳论文奖（2016年、2022年）、两次获得 SIGMOD 最佳论文荣誉提名（2022年、2024年），并获得 SIGMOD 最佳演示奖（2015年）和 PODS 时间检验奖（2022年）。