A Platform-Free Proof of Federated Learning Consensus Mechanism for Sustainable Blockchains
Yuntao Wang,Haixia Peng,Zhou Su,Tom H. Luan,Abderrahim Benslimane,Yuan Wu
DOI: https://doi.org/10.1109/jsac.2022.3213347
IF: 16.4
2022-11-26
IEEE Journal on Selected Areas in Communications
Abstract:Proof of work (PoW), as the representative consensus protocol for blockchain, consumes enormous amounts of computation and energy to determine bookkeeping rights among miners but does not achieve any practical purposes. To address the drawback of PoW, we propose a novel energy-recycling consensus mechanism named platform-free proof of federated learning (PF-PoFL), which leverages the computing power originally wasted in solving hard but meaningless PoW puzzles to conduct practical federated learning (FL) tasks. Nevertheless, potential security threats and efficiency concerns may occur due to the untrusted environment and miners' self-interested features. In this paper, by devising a novel block structure, new transaction types, and credit-based incentives, PF-PoFL allows efficient artificial intelligence (AI) task outsourcing, federated mining, model evaluation, and reward distribution in a fully decentralized manner, while resisting spoofing and Sybil attacks. Besides, PF-PoFL equips with a user-level differential privacy mechanism for miners to prevent implicit privacy leakage in training FL models. Furthermore, by considering dynamic miner characteristics (e.g., training samples, non-IID degree, and network delay) under diverse FL tasks, a federation formation game-based mechanism is presented to distributively form the optimized disjoint miner partition structure with Nash-stable convergence. Extensive simulations validate the efficiency and effectiveness of PF-PoFL.
telecommunications,engineering, electrical & electronic