Detecting model misconducts in decentralized healthcare federated learning

Area of Science:

• Genomic Medicine
• Artificial Intelligence
• Machine Learning

Background:

• Cross-institutional collaborations using AI on clinical/genomic data accelerate research.
• Risks of incorrect model submission (accidents or malicious intent) can deter participation in federated learning.
• Existing methods for model misconduct are algorithm-specific.

Purpose of the Study:

• To develop an algorithm-agnostic framework for detecting model misconduct in federated learning.
• To address the challenge of ensuring integrity and reliability in collaborative AI research.

Main Methods:

• Designed a simulator to generate various types of model misconduct (Plagiarism, Fabrication, Falsification).
• Developed a detection framework with Auditing, Coefficient, and Performance detectors.
• Employed greedy parameter tuning for optimal detection.

Main Results:

• Evaluated the detection method on three datasets, generating 10 misconduct types.
• Achieved high recall with low additional computational cost.
• Successfully identified misconduct on specific sites across learning iterations, though precise site/iteration detection remains challenging.

Conclusions:

• The proposed method enhances the integrity and reliability of federated machine learning for genomic and healthcare data.
• Supports trustworthy collaborative AI research in sensitive data domains.