Speaker: Dr. Sinan Yıldırım (Sabancı University)
Title: Sequentially Interactive Learning with Local Differential Privacy
Date: 17 October 2025, Friday
Time: 13:30-14:30
Place: EA409
Abstract: In privacy-preserving data analysis, sensitive personal data are collected after being privatized via randomization. This talk concerns online parameter estimation under data privacy constraints. In particular, local differential privacy (LDP) will be considered the definition of data privacy. The main research question of this study is how to estimate a population distribution efficiently under LDP. The key idea behind our methodology is to use a “tunable” LDP mechanism, which is tuned via sequential interaction: At each iteration, the method goes over the following steps: (1) estimates the parameter from the privatized data points collected so far, (2) tunes an adaptable LDP mechanism to optimize the estimation utility of future privatized data, and (3) collects the next data point using the adapted LDP mechanism. Step (2) is a constraint optimization problem and, in our method, it is generally solved suboptimally by confining the search space to parametric families of LDP mechanisms. Numerical studies on a variety of scenarios demonstrate the merits of our methodology. (Ongoing work with my MSc student Giray Düzel.)
Bio: Sinan Yıldırım, PhD, has been a faculty member since 2015 in the Faculty of Engineering and Natural Sciences at Sabancı University, Turkey. He received his BS in 2007 and MS in 2009, both in Electrical and Electronics Engineering at Boğaziçi University, Turkey. He holds a PhD in Mathematical Statistics from the University of Cambridge, UK, where he graduated in 2013. He then worked as a postdoctoral researcher at the University of Bristol from 2013 to 2015. His primary research areas are Bayesian Statistics, Monte Carlo methods, and data privacy.