BlockDFL: A Blockchain-based Fully Decentralized Peer-to-Peer Federated Learning Framework
BlockDFL leverages blockchain to enable fully decentralized P2P federated learning, resisting poisoning, enhancing privacy, and reducing communication overhead.
Key Findings
Methodology
This framework integrates PBFT-based voting and a two-layer scoring mechanism, using blockchain to assign roles randomly and achieve consensus without trust. Gradient compression reduces communication load and prevents data reconstruction. The local updates are filtered via median testing, and global updates are scored with Krum to resist poisoning. Roles include update providers, aggregators, and verifiers, ensuring secure, private, and scalable training. Experiments on MNIST and CIFAR-10 demonstrate robustness against 40% malicious participants, maintaining accuracy comparable to centralized FL, with over 85% communication cost reduction.
Key Results
- On MNIST, BlockDFL maintains about 98% accuracy with 40% malicious nodes, outperforming existing blockchain-based P2P FL frameworks, with communication savings exceeding 85%.
- On CIFAR-10, gradient sparsification boosts accuracy to 75%, with a 3% improvement over uncompressed models, and enhances poisoning resistance, confirming the effectiveness of the sparsity strategy.
- Results show stable performance in non-IID settings, with role randomness and two-layer scoring effectively defending against poisoning, validating the system's security and fairness.
Significance
This work advances blockchain-based decentralized federated learning, addressing trust, security, and privacy issues. It offers a scalable, efficient solution suitable for edge and IoT environments, promoting trustworthy AI deployment. The integration of PBFT consensus and multi-layer filtering sets a new standard for secure distributed AI, bridging gaps between theory and real-world applications.
Technical Contribution
Key innovations include a PBFT-inspired voting mechanism ensuring single global update per round, a two-layer scoring system for robust poisoning defense, gradient compression for privacy and efficiency, and role randomization to prevent collusion. These components collectively create a high-performance, trustworthy decentralized FL framework that surpasses prior solutions in security, scalability, and privacy protection.
Novelty
This is the first work combining PBFT consensus with blockchain-based P2P FL, employing a dual scoring strategy to effectively counter poisoning attacks. Unlike prior methods relying on centralized trust or simple aggregation, it achieves a fully trustless, scalable environment with strong security guarantees, filling a crucial gap in decentralized AI research.
Limitations
- Role assignment randomness may still be vulnerable under extreme adversarial conditions, requiring further role scheduling optimization.
- Gradient sparsification, while reducing communication, may impact accuracy in highly non-IID data scenarios, necessitating better trade-offs.
- Large-scale deployment and real-world blockchain integration pose challenges in terms of latency and storage, which need further exploration.
Future Work
Future efforts will focus on adaptive role allocation, integrating differential privacy and homomorphic encryption for enhanced privacy, optimizing consensus protocols for massive networks, and extending the framework to heterogeneous devices, aiming for practical, large-scale deployment.
AI Executive Summary
Distributed AI has become a cornerstone of modern technology, but its widespread adoption faces significant hurdles in trust, privacy, and efficiency. Traditional federated learning (FL) relies heavily on centralized servers, which pose single points of failure and trust issues. To overcome these limitations, this paper introduces BlockDFL, a blockchain-based fully decentralized P2P federated learning framework.
BlockDFL leverages blockchain's inherent decentralization, employing a PBFT-inspired voting mechanism to ensure model update consensus without forks. The system introduces a two-layer scoring process: local updates are filtered through median testing, and global updates are scored with Krum, effectively filtering poisoned updates. Gradient compression reduces communication overhead and enhances data privacy, especially in non-IID data environments.
Extensive experiments on MNIST and CIFAR-10 datasets demonstrate that BlockDFL maintains high accuracy even with 40% malicious participants, outperforming existing blockchain-based P2P FL frameworks. Communication costs are reduced by over 85%, and the system exhibits strong scalability and robustness against poisoning attacks. These results highlight the potential of blockchain to enable secure, trustworthy, and efficient decentralized AI.
Despite promising outcomes, challenges remain in optimizing role assignment under adversarial conditions, balancing sparsity with accuracy, and scaling to large networks. Future research will explore adaptive role strategies, privacy-preserving cryptographic techniques, and more efficient consensus algorithms. Overall, this work paves the way for practical, secure, and scalable decentralized AI systems, with broad implications for edge computing, IoT, and beyond.
Deep Dive
Abstract
Federated learning (FL) enables collaborative training of machine learning models without sharing training data. Traditional FL heavily relies on a trusted centralized server. Although decentralized FL eliminates the central dependence, it may worsen the other inherit problems faced by FL such as poisoning attacks and data representation leakage due to insufficient restrictions on the behavior of participants, and heavy communication cost, especially in fully decentralized scenarios, i.e., peer-to-peer (P2P) settings. In this paper, we propose a blockchain-based fully decentralized P2P framework for FL, called BlockDFL. It takes blockchain as the foundation, leveraging the proposed PBFT-based voting mechanism and two-layer scoring mechanism to coordinate FL among peer participants without mutual trust, while effectively defending against poisoning attacks. Gradient compression is introduced to lowering communication cost and prevent data from being reconstructed from transmitted model updates. Extensive experiments conducted on two real-world datasets exhibit that BlockDFL obtains competitive accuracy compared to centralized FL and can defend poisoning attacks while achieving efficiency and scalability. Especially when the proportion of malicious participants is as high as 40%, BlockDFL can still preserve the accuracy of FL, outperforming existing fully decentralized P2P FL frameworks based on blockchain.