BlockDFL: A Blockchain-based Fully Decentralized Peer-to-Peer Federated Learning Framework

TL;DR

BlockDFL presents a blockchain-based fully decentralized P2P federated learning framework that defends against poisoning attacks and data leakage while maintaining high efficiency. It combines a PBFT-based voting mechanism with a two-layer scoring system (median-based testing for local updates and Krum for global updates) and employs gradient compression to reduce communication and obscure training data. The system assigns roles via stake-weighted hash rings and uses a four-phase round (role selection, local training, aggregation, verification/consensus) to update a single global model per block without forking. Empirical results on MNIST and CIFAR-10 show competitive accuracy with centralized FL and strong poisoning tolerance up to 40% malicious participants, along with favorable efficiency and scalability relative to existing fully decentralized blockchain-based FL frameworks.

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.

Figures (9)

• Figure 1: The detailed processes of BlockDFL in one round of communication, from ➀ to ➃.
• Figure 2: Average test accuracy and the corresponding standard deviation of approaches in the last 20% rounds.
• Figure 3: Successful attack ratio in approaches with different percentage of malicious participants.
• Figure 4: Convergence of the global model test accuracy obtained by vanilla FL and BlockDFL.
• Figure 5: Time consumption of processes in BlockDFL with varying number of participants, aggregators and verifiers.