Hung Do
Title: Differentially Private Stochastic Block Models
Date: Thursday, August 20th, 2026
Time: 2:00pm
Location: LIB 2020 & Zoom
Supervised by: Owen Ward
Abstract:
Graph data often encodes sensitive relationships between individuals, making privacy-preserving analysis essential when such data are shared or published. This thesis introduces a differentially private inference framework for the Stochastic Block Model (SBM), a widely used model for community structure in networks. We derive a mean-field variational approximation to the SBM posterior and fit it using Differentially Private Stochastic Gradient Descent (DP-SGD), enabling scalable, mini-batch training directly on the Evidence Lower Bound. Our method infers both node community labels and block-wise edge probabilities simultaneously under formal $(\epsilon, \delta)$-differential privacy guarantees, and is extendable to alternative edge distributions and local differential privacy.