Federated influencer learning for secure and efficient collaborative learning in realistic medical database...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep learning requires large datasets, but centralized training poses data ownership and security risks.
• Traditional federated learning (FL) has limitations including vulnerability to attacks, non-IID data, central server reliance, and communication overhead.
• FL is not optimized for dynamic hospital database environments.

Purpose of the Study:

• To introduce Federated Influencer Learning (FIL) as a secure and efficient collaborative learning paradigm.
• To address the limitations of FL in resource-constrained and privacy-sensitive environments like hospitals.
• To develop a model-agnostic training approach suitable for dynamic data accumulation.

Main Methods:

• FIL employs an equal-status structure with an administrator, unlike FL's server-client model.
• The FIL process includes four stages: local training, qualification (influencers/followers), screening (logit integrity), and influencing (knowledge sharing).
• FIL avoids model-parameter transactions and central servers, enhancing security and enabling model-agnostic training.

Main Results:

• FIL demonstrated superior performance compared to several FL methods on medical (X-ray, MRI, PET) and natural (CIFAR-10) image datasets.
• Experiments were conducted in a dynamically accumulating database environment, showcasing FIL's adaptability.
• FIL achieved higher precision, recall, and Dice scores, with lower standard deviation across participants, notably a 40% improvement in Dice score and recall for PET data.

Conclusions:

• FIL offers enhanced security and efficiency over traditional FL by eliminating central servers and model-parameter exchange.
• The model-agnostic nature and equal-status structure make FIL highly suitable for healthcare and other privacy-critical fields.
• FIL effectively addresses the challenges of collaborative learning in dynamic, distributed environments.