A Cost-Aware Utility-Maximizing Bidding Strategy for Auction-Based Federated Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Auction-based federated learning (AFL) incentivizes data contribution.
• Existing methods overlook the generalized second-price (GSP) auction's cost mechanism.
• Data consumers (DCs) need effective bidding strategies in competitive AFL markets.

Purpose of the Study:

• To propose a novel federated cost-aware bidding strategy for DCs in GSP auction-based FL.
• To enable DCs to maximize their utility and improve FL model accuracy.
• To address the open problem of optimal bidding for data acquisition in AFL.

Main Methods:

• Formulated the optimal bidding function under GSP auction rules.
• Developed a framework that jointly optimizes utility estimation and market price modeling.
• Implemented a return on investment (ROI)-based method for optimal bid price determination.

Main Results:

• The proposed strategy significantly outperforms eight state-of-the-art methods.
• Achieved average improvements in data acquisition, cost-efficiency, utility, and FL model accuracy.
• Demonstrated superior performance across six benchmark datasets.

Conclusions:

• The federated cost-aware bidding strategy effectively maximizes DC utility in GSP-based AFL.
• The approach enhances overall FL model performance and data acquisition efficiency.
• This work provides a crucial solution for competitive AFL market dynamics.