MSc Thesis Defense: Giray Düzel, On the Use of Online Estimation for Local Differential Privacy with Sequential Interaction, Date & Time: June 25, 2026 – 1:30 PM, Place: FENS 2019

On the Use of Online Estimation for Local Differential Privacy with Sequential Interaction

Giray Düzel
Data Science, MSc Thesis, 2026

Thesis Jury

     Assoc. Prof. Sinan Yıldırım (Thesis Advisor)

  Prof. Albert Levi

  Assoc. Prof. Mehmet Emre Gürsoy

Date & Time: June 25th, 2026 –  1:30 PM

Place: FENS 2019

Keywords : local differential privacy, frequency estimation, sequential interactivity, adaptive mechanisms, online estimation

Abstract

Discrete distribution estimation under local differential privacy is a fundamental task in privacy-preserving data analytics. However, existing differential privacy mechanisms are commonly static, ignoring the population distribution and thus losing utility, or require separate collection phases. This thesis develops a sequentially interactive framework in which the privatization mechanism is adapted online to a maintained estimate while preserving the same local privacy guarantee for every participant. It introduces estimate-aware hashing and hybrid hashing-subset selection mechanisms, together with scalable moment-based, maximum-likelihood, and Bayesian estimators for online and offline inference. A constrained-optimization analysis characterizes the globally optimal locally private channel under the probability of honest response utility. The resulting active-set mechanism is shown to sacrifice tail identifiability, and some strategies are proposed to alleviate this effect. Synthetic and real-data experiments demonstrate that adaptive mechanisms generally improve estimation utility over static baselines, especially for concentrated distributions.