SecMDGM: Federated Learning Security Mechanism Based on Multi-Dimensional Auctions.

Area of Science:

• Distributed Machine Learning
• Information Security
• Auction Theory

Background:

• Federated learning (FL) is an emerging distributed machine learning technology with significant advantages in the big data era.
• Traditional auction theory inadequately addresses quality considerations and privacy concerns in multi-dimensional auctions.
• Existing mechanisms lack the necessary security and effectiveness for motivating active and safe participant engagement in FL.

Purpose of the Study:

• To develop a secure mechanism for motivating participants in federated learning auctions.
• To ensure high-quality data or computational performance from winning participants.
• To address the challenges of multi-dimensional auctions considering both price and quality, alongside privacy.

Main Methods:

• Proposed a novel multi-dimensional information security mechanism.
• Developed an optimal mechanism, named SecMDGM, satisfying Pareto optimality and incentive compatibility.
• Conducted theoretical proofs and experimental verification to validate the mechanism's effectiveness.

Main Results:

• The proposed SecMDGM mechanism enhances participant motivation and ensures data quality in federated learning auctions.
• Theoretical and experimental validation confirmed the mechanism's security, necessity, and effectiveness.
• Demonstrated a 2.73-fold performance improvement for aggregation models based on vertical data compared to random selection.

Conclusions:

• The SecMDGM mechanism offers a practical and significant advancement for secure multi-dimensional auctions in federated learning.
• This mechanism effectively addresses privacy concerns and improves data/computational performance.
• The findings have practical implications for implementing secure and efficient federated learning systems.