Computer Science Faculty Publications

Lightweight Privacy-First Federated Learning for Medical AI

Document Type

Conference Proceeding

Publication Date

7-20-2026

Abstract

Medical AI systems require large, diverse datasets from multiple institutions to achieve clinical-grade diagnostic accuracy. However, aggregating patient data into centralized repositories violates privacy regulations. To solve this, researchers developed federated learning (FL) that lets hospitals keep their data private while still training AI together, but they still transmit weights/gradients to the central server, which introduces new privacy attack surfaces and computational overhead. These challenges intensify when inputs are privacy-transformed (e.g., block-wise permutations or sign flips), where conventional convolutional networks (CNNs) lose critical spatial structure. To address these issues, this study proposes a lightweight, privacy-first federated framework that shares only compact [CLS] embeddings from client-side Vision Transformers (ViTs), never transmitting weights, gradients, or pixels. ViTs’ global self-attention still captures key patterns even when images are patched or shuffled, maintaining accuracy on obfuscated inputs. Adding Gaussian DP noise to the shared embeddings further reduces inference and inversion risks with minimal accuracy loss. On colon histopathology data, our method achieves 96% validation accuracy while cutting communication by about 20 times compared to gradient sharing, making it practical for real-world medical AI with strong privacy and low bandwidth needs.

Publication Title

nternational Conference on Security and Privacy in Communication Systems

DOI

10.1007/978-3-032-32764-2_13

Link to Full Text

Share

COinS