Federated Learning: Issues in Medical Application

Joo Hun Yoo, Hyejun Jeong, Jaehyeok Lee, Tai-Myoung Chung

Abstract

Since federated learning (FL), which enables AI training without moving local data, was introduced by Google in 2017, it has been actively studied, particularly in the field of medicine. The concept of machine learning without collecting data from local clients is highly attractive because sensitive data remains on-site. However, FL techniques still face various open issues due to characteristics such as non-identical data distribution, client participation management, and vulnerable environments. This paper provides an overview of current issues—including data/system heterogeneity, client management, traceability, and security—to make FL flawlessly useful in the real world. Additionally, we introduce a modularized FL framework currently under development to experiment with various techniques and protocols to find solutions for these issues.

Introduction

Medical AI requires massive amounts of data to train accurate predictive models. However, strict privacy laws (like GDPR, CCPA, and HIPAA) restrict the sharing of Personal Health Information (PHI). Federated Learning solves this by bringing the model to the data: local clients (hospitals or wearable devices) train a model on their private data and only share the updated model weights with a central server. While FL works well in theory, deploying it in real-world medical settings involves significant hurdles.

Critical Challenges in Medical FL

Transitioning FL from research to clinical practice reveals four primary gaps that standard algorithms often fail to address:

  1. Heterogeneity: Hospitals have vastly different patient demographics (data heterogeneity) and IT infrastructure (system heterogeneity). This "Non-IID" (Non-Independent and Identically Distributed) nature of medical data makes it difficult for global models to converge.
  2. Client Management: There is a lack of mechanisms to incentivize hospitals with high-quality data to participate or to prevent "free-riders" from reaping benefits without contributing fairly.
  3. Traceability: The privacy-preserving nature of FL creates an accountability "black-box." If a model makes a false-positive diagnosis, the inability to inspect raw training data makes it nearly impossible to audit the system.
  4. Security: Because the server cannot inspect local data, the system is vulnerable to malicious actors injecting bad data (poisoning) or trying to reverse-engineer patient data from model weights (inference attacks).
Overview of Federated Learning architectures at the hospital and patient levels.

Methodological Approaches and Client Management

To address these gaps, researchers are exploring several advanced techniques reviewed in this survey:

  • Handling Heterogeneity: Approaches include clustering hospitals with similar data distributions and using adaptive optimization algorithms (like FedProx or SCAFFOLD) to prevent local models from drifting too far from the global objective.
  • Security Defenses: Techniques like Differential Privacy and Multi-Party Computation (MPC) are used to thwart inference attacks, while robust aggregation methods help detect and ignore suspicious model updates.

Incentive mechanisms are also critical. Using game theory (e.g., Shapley Values) and auction mechanisms, a central server can evaluate a client's computational and data contributions, rewarding them accordingly to ensure long-term sustainability of the network.

The process of client evaluation and incentive distribution within an FL environment.

The Modular FL Framework

A significant contribution of this work is the introduction of a Modular FL framework. Designed to be lightweight and flexible, this framework separates the FL architecture into distinct modules: global model distribution, client selection, secure aggregation, and serving. By saving model weights as JSON files, the framework remains language-agnostic, allowing researchers to easily swap out specific functions—such as testing a new incentive protocol—without re-engineering the entire system. This framework will be open-sourced upon completion.

Conclusion and Future Work

While FL consistently achieves predictive accuracy comparable to centralized learning across various medical tasks, the "Accountability Trade-off" remains a significant hurdle. Future research must prioritize the integration of Explainable AI (XAI) with FL (e.g., using techniques like Grad-CAM) to ensure fail-safe, auditable usage in clinical settings. The upcoming Modular FL framework aims to provide a much-needed testbed for researchers to tackle these exact issues.

Summary updated

Abstract

Since the federated learning, which makes AI learning possible without moving local data around, was introduced by google in 2017 it has been actively studied particularly in the field of medicine. In fact, the idea of machine learning in AI without collecting data from local clients is very attractive because data remain in local sites. However, federated learning techniques still have various open issues due to its own characteristics such as non identical distribution, client participation management, and vulnerable environments. In this presentation, the current issues to make federated learning flawlessly useful in the real world will be briefly overviewed. They are related to data/system heterogeneity, client management, traceability, and security. Also, we introduce the modularized federated learning framework, we currently develop, to experiment various techniques and protocols to find solutions for aforementioned issues. The framework will be open to public after development completes.
BibTeX
@inproceedings{yoo2021federated,
  author = {Joo Hun Yoo and Hyejun Jeong and Jaehyeok Lee and Tai-Myoung Chung},
  title = {{Federated Learning: Issues in Medical Application}},
  booktitle = {FDSE},
  year = {2021},
  url = {https://openreview.net/forum?id=mqs4kcaij0M},
}
Figure