BFLP: blockchain-based federated learning protocol with verifiable contribution proof

Abstract

Cross-institutional federated learning suffers from centralized trust fragility, poisoning attacks, and the lack of verifiable contribution-based incentives. This paper proposes BFLP, a blockchain-based federated learning protocol integrating zero-knowledge proofs (ZKPs) into a layered, contract-driven architecture. Its core innovation is a dual-proof mechanism that provides cryptographic evidence of procedural compliance: a data quality proof (πquality\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi _{\text {quality}}$$\end{document}) certifies that local data meets a public quality threshold, while a training correctness proof (πtraining\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi _{\text {training}}$$\end{document}) binds the submitted update to the committed data and prescribed training algorithm. This mechanism guarantees computational correctness and update integrity, but not the utility of the update for the global model; contribution utility is separately assessed by the incentive layer. This design is feasible only for lightweight models with up to a few hundred thousand parameters and deterministic training pipelines; it constitutes a proof-of-concept for verifiable compliance in blockchain-based FL rather than a ready-to-deploy protocol for production-scale deep learning. Smart contracts automate task management, on-chain proof verification, secure aggregation, and fair reward distribution. Experiments with up to 100 nodes on Fashion-MNIST and CIFAR-10 using a lightweight CNN show a throughput of 785 TPS; under 20% malicious nodes, model accuracy remains at 87.1% (Fashion-MNIST), the malicious detection rate reaches 96.8%, and ZKP verification latency is below 0.8 seconds. These results indicate that BFLP can enable secure, verifiable, and fair federated collaboration with manageable overhead on modern image benchmarks.