Shortest path core-selection incentive for federated learning platform with medical applications

Area of Science:

• Computer Science
• Artificial Intelligence
• Game Theory

Background:

• Federated learning (FL) is crucial for addressing data islands and extracting value from distributed data.
• Emerging FL platforms organize users, devices, and data for collaborative tasks via crowdsourcing.
• Existing mechanisms face challenges like overpayment and security vulnerabilities.

Purpose of the Study:

• To design an incentive mechanism for federated learning platforms that improves upon the VCG auction.
• To address overpayment, false-name attacks, and core deviation issues inherent in the VCG mechanism.
• To ensure fair and efficient rewards for federated learning task participants.

Main Methods:

• Combines the VCG (Vickrey-Clarke-Groves) auction mechanism with cooperative game theory's core concept.
• Introduces a shortest path core-selection strategy for participant reward allocation.
• Applies the VCG-nearest principle for selecting participants and determining rewards.

Main Results:

• The proposed mechanism guarantees incentive compatibility, individual rationality, and alliance rationality.
• It ensures maximization of social efficiency within the federated learning task.
• Medical experimental results validate the mechanism's effectiveness in real-world scenarios.

Conclusions:

• The shortest path core-selection incentive mechanism offers a robust solution for federated learning platforms.
• It enhances economic attributes and reduces expenditure for task issuers.
• The mechanism provides a fairer reward distribution, closely approximating VCG outcomes.