Project Description
Medical studies often require using sensitive individual patient data. Survival models, e.g., Cox’s Proportional Hazard Model, are among the most popular statistical model for medical studies involving estimating personalized survival time for terminal diseases. More broadly, they are also used to estimate the time of marriage and divorce, the time of machine failure in data centers, and the time of a website getting hacked, among a broader range of applications beyond the medical context. The training (or parameter fitting) of the survival models, however, requires the use of sensitive individual data, which may cause the individual patients to be re-identified. This project develops new techniques for training survival models with strong differential privacy guarantees — the gold standard for protecting individual privacy. The goal is to achieve the best privacy-utility tradeoff by leveraging various new techniques developed in modern differentially privacy accounting and deep learning with DP
Team Members
- Elaine Ho
- Adyah Rastogi
- Oscar Benedeck
- Philly Lim
Professor and Mentors
- Prof. Yu-Xiang Wang
- Grad mentor: Erchi Wang and Esha Singh
Meeting Time
- Meeting with the Professor
- Fridays at 4p
- Meeting with Grad mentor
- TBD
- ERSP meeting with central mentors
- Chinmay: TBD
- Diba: TBD
- ERSP team meetings
- Thursdays 4-6p
Links to Proposals and Presentation
- Instructor feedback on initial draft of proposal: link
- Final Proposal (after instructor's feedback): link
- Final presentation: link