Blockchain-Based Federated Learning System: A Survey on Design Choices

Area of Science:

• Artificial Intelligence
• Computer Science
• Cybersecurity

Background:

• Federated learning (FL) traditionally assumes trusted environments, limiting its application in real-world scenarios requiring collaboration among untrusted parties.
• Blockchain technology offers a decentralized and immutable platform, making it a promising solution for enhancing trust in federated learning systems.
• Recent research interest has focused on integrating blockchain with federated learning to address security and trust concerns.

Purpose of the Study:

• To conduct a comprehensive literature survey of state-of-the-art blockchain-based federated learning (BFL) systems.
• To analyze common design patterns and variations used to overcome challenges in BFL.
• To evaluate the pros and cons of different design choices based on key performance metrics.

Main Methods:

• Systematic literature review of existing BFL research.
• Identification and categorization of approximately 31 distinct design variations across BFL systems.
• Analysis of design patterns considering robustness, efficiency, privacy, and fairness.

Main Results:

• A linear relationship was observed between fairness and robustness, indicating that improvements in fairness can enhance robustness.
• A trade-off exists between improving all metrics simultaneously and overall system efficiency.
• Popular design choices and areas needing further development in BFL were identified.

Conclusions:

• Future BFL systems require advancements in model compression and asynchronous aggregation techniques.
• There is a need for more rigorous system efficiency evaluations in BFL research.
• Further research should focus on applying BFL to cross-device settings and addressing identified efficiency trade-offs.