Faculty Research

Overview

Modern vision systems (healthcare, surveillance, remote sensing) operate in distributed environments where data is siloed across institutions. Centralized learning is often infeasible due to privacy, ownership, and regulatory constraints.

This research develops a unified federated learning framework that enables collaborative visual recognition while addressing key real-world challenges:

• Non-IID data
• Model and label heterogeneity
• Privacy risks in visual data
• Lack of semantic understanding

Core Idea

Instead of sharing raw images, clients train locally and share only compact knowledge (representations, statistics, or predictions). A central server intelligently aggregates this information to build a robust global model.

Key Challenges Addressed

• High-dimensional data → redundancy, unstable learning
• Non-IID distributions → slow and divergent convergence
• Heterogeneous clients → incompatible models and labels
• Privacy leakage → sensitive visual content exposure
• Semantic gap → purely visual features lack meaning