Dynamic Trust Modeling in Federated Learning Through Balancing Utility and Privacy

The surge in data-intensive machine learning (ML) applications necessitates effective incentives for data owners (DOs) to contribute data and train ML models collaboratively. The decision to participate in collaboration depends on the balance between utility gains and privacy loss. This project focuses on federated learning (FL), where DOs participate in collaborative learning without sharing raw data. While FL preserves privacy to some extent, vulnerabilities exist in preserving data privacy through shared models. Existing literature proposes privacy guarantees but lacks a clear method to measure the utility gain of each participant. The project aims to study participants’ contributions to collaborative learning and determine compensation based on different privacy levels. The project then investigates the utility gain of each participant. In practice, FL often faces data heterogeneity, resulting in different privacy needs. This can prompt DOs to exit at different times, managing privacy budgets according to their gained utility. Exits can deplete the data pool, impoverishing model quality and potentially triggering a cascade effect. Conversely, rapid disengagement by some participants may inspire others to contribute more data, aiding joint training for a personalized model. The project explores these dynamic utility-privacy trade-offs, analyzing participants’ evolving trust in the FL procedure.

Faculty Supervisor:

Stark C. Draper

Student:

Partner:

CISPA

Discipline:

Engineering

Sector:

Artificial Intelligence; Information and Communications Technology

University:

University of Toronto

Program: