FedOSS: Federated Open Set Recognition via Inter-Client Discrepancy and Collaboration.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Open set recognition (OSR) in medicine aims to classify known diseases and identify unknown ones.
• Traditional OSR faces privacy and security risks due to centralized data aggregation.
• Federated learning (FL) offers a privacy-preserving solution for distributed medical data.

Purpose of the Study:

• To introduce Federated Open Set Recognition (FedOSR) as a novel approach for medical AI.
• To propose the Federated Open Set Synthesis (FedOSS) framework to address the challenge of unavailable unknown samples in FL.
• To enhance the ability of medical AI to distinguish between known diseases and novel, unseen conditions.

Main Methods:

• Developed the Federated Open Set Synthesis (FedOSS) framework for FedOSR.
• Introduced Discrete Unknown Sample Synthesis (DUSS) to generate virtual unknown samples using inter-client knowledge.
• Implemented Federated Open Space Sampling (FOSS) to estimate open data space distributions and improve sample diversity.

Main Results:

• FedOSS effectively generates virtual unknown samples crucial for learning decision boundaries.
• DUSS and FOSS modules demonstrated significant contributions to the framework's performance through ablation studies.
• The FedOSS framework achieved superior performance compared to existing state-of-the-art methods on medical datasets.

Conclusions:

• FedOSS provides an effective solution for privacy-preserving open set recognition in medical AI.
• The framework successfully tackles the challenge of unknown sample scarcity in federated learning settings.
• This work represents a significant advancement in applying federated learning to complex medical diagnostic tasks.