A Blockchain-Based Robust Federated Learning Framework and Proof-of-Learning Contribution (PoLC) Consensus

다국어 초록 (Multilingual Abstract)

Federated Learning (FL) has emerged as a prominent alternative for addressing data privacy issues in artificial intelligence training; however, its reliance on a central coordination server for aggregation leads to Single Point of Failure (SPoF) and reliability concerns. Although Blockchain-based Federated Learning (BFL) has been proposed to overcome these issues, existing research faces fundamental limitations, such as blockchain scalability constraints, state bloat, and the incentive misalignment between capital (stake) and learning contribution resulting from the direct use of Proof-of-Stake (PoS) governance.
To address the complex issues inherent in existing FL and BFL, this paper proposes D-FLIB (Decentralized Federated Learning via Inter-Blockchain Communication), a novel decentralized federated learning framework. D-FLIB is based on two core designs. First, it employs a hierarchical dual-chain architecture grounded in the principle of separation of concerns. This architecture separates the coordination layer chain (Fedlearning), responsible for the actual execution of federated learning, aggregation tasks, and governance, from the data layer chain (Fedstoraging), designed for data-intensive operations such as storage, availability, and access control. These chains interoperate via Inter-Blockchain Communication (IBC) in a trust-minimized manner to maximize scalability and efficiency. Second, we propose a novel consensus protocol, Proof-of-Learning-Contribution (PoLC), to fundamentally address the governance issues of BFL. PoLC replaces the economic stake of PoS with on-chain verified, quantitative learning contributions (ATT/EWMA scores). This ensures that governance authority within the blockchain system is proportional to actual contributions and establishes a dynamic value-based governance structure that aligns the incentives of learning participants with the system's long-term objective of generating high-quality models. Furthermore, D-FLIB securely controls access to sensitive data between the two independent chains without reliance on a trusted entity, utilizing a Cross-Chain Role-Based Access Control (Cross-Chain RBAC) mechanism powered by IBC Interchain Queries (ICQ).
The proposed D-FLIB framework was prototyped based on the Cosmos Core, and simulations were conducted using the CIFAR-10 and CIFAR-100 datasets in a Non-IID data environment with a Dirichlet distribution (α=0.5). In simulation environments, D-FLIB achieved a superior AI model accuracy of 79.66% compared to standard FedAvg (77.12%); furthermore, even in extreme scenarios involving 50% data poisoning attackers, it maintained a stable accuracy of 72.09%—a 14.84% improvement over the general FL model (57.25%)—demonstrating stable accuracy convergence. This experimentally demonstrates that the PoLC consensus functions as an intrinsic defense mechanism that identifies malicious learning nodes and blocks their influence within the network.
By presenting a specialized blockchain framework that integrates a scalable architecture for federated learning, a novel value-based governance model (PoLC), and a modular security pattern (Cross-Chain RBAC), this study provides a core foundation for a practical and robust decentralized AI ecosystem, marking a significant academic and technical contribution.

목차 (Table of Contents)

• 1. Introduction 1
• 1.1. Research Objectives 2
• 1.2. Proposed Methodology: D-FLIB Framework and PoLC Consensus 5
• 1.3. Research Contribution 6
• 2. Theoretical Background & Related Work 9
• 2.1. Federated Learning 9
• 2.1.1. Aggregation in Federated Learning: FedAvg Algorithm 10
• 2.1.2. Key Challenges in Federated Learning 12
• 2.1.3. Research on Contribution Evaluation in Federated Learning 14
• 2.2. Distributed Systems and Blockchain 15
• 2.2.1. Application-Specific Blockchain Paradigm 15
• 2.2.2. Inter-Blockchain Commnuication 17
• 2.3. Decentralized Storage 19
• 2.3.1. Related Technologies 19
• 2.3.2. Data Durability 20
• 2.4. Prior Research on Blockchain-based Federated Learning 21
• 2.4.1. Type 1: L1 Smart Contract-Based Models 22
• 2.4.2. Type 2: Hybrid Storage Models 23
• 2.4.3. Type 3: Single App-Chain Models 24
• 2.4.4. Type 4: L2 (Rollup) Utilization Models 24
• 2.4.5. Research Gap 25
• 3. D-FLIB: PoLC-Based Decentralized Federated Learning Protocol 27
• 3.1. PoLC Protocol: Architectural Philosophy 27
• 3.2. System Model 29
• 3.2.1. Participant Model 29
• 3.2.2. Network and Security Threat Model 30
• 3.3. D-FLIB Dual-Chain Architecture Design 32
• 3.3.1. Coordination Layer Chain (fedlearning): Coordination and Governance 34
• 3.3.2. Data Layer Chain (fedstoraging): Data Availability and Access Control 36
• 3.3.3. IBC: Inter-Blockchain Communication 38
• 3.4. Coordination Layer Chain (fedlearning) Protocol Design 40
• 3.4.1. State-Based Round Protocol (FSM) 40
• 3.4.2. On-Chain Contribution Proof and Aggregation 43
• 3.4.3. Dynamic Value-Based Governance 49
• 3.4.4. PoLC: Contribution-Based Consensus 51
• 3.5. Data Layer Chain (fedstoraging) Protocol Design 54
• 3.5.1. Proof-of-Retrievability (PoR) 55
• 3.5.2. Authenticated Data Access Mechanism(Cross-Chain RBAC) 58
• 4. Implementation 64
• 4.1. Development Environment and Technology Stack 64
• 4.1.1. Blockchain Development Environment 65
• 4.1.2. Federated Learning Development Environment 66
• 4.2. D-FLIB Component Implementation 67
• 4.2.1. fedlearning Chain 67
• 4.2.2. fedstoraging Chain 68
• 4.2.3. IBC-Relayer 70
• 4.3. Federated Learning Component Implementation 72
• 4.3.1. Training Dataset and Distributed Environment 72
• 4.3.2. Learning Model Architecture 74
• 4.3.3. Federated Learning Protocol and Client Implementation 75
• 5. Experiments and Performance Evaluation 77
• 5.1. Experimental Environment Setup 77
• 5.2. Protocol Functional Verification 82
• 5.2.1. FSM-based Round State Transition Verification 82
• 5.2.2. On-Chain Contribution Aggregation (ATT) and Dynamic Governance Verification 85
• 5.3. Performance Benchmarking 89
• 5.3.1. Measurement of On-Chain and Cross-Chain Processing Latency 89
• 5.3.2. Network Load Measurement 93
• 5.3.3. Federated Learning Performance Measurement 94
• 5.4. Security and Robustness Verification 96
• 5.4.1. Data Poisoning and Sybil Attack Scenarios 96
• 5.4.2. Random Noise Attack Scenario 104
• 6. Discussion 107
• 6.1. Implications of D-FLIB and PoLC Protocol 107
• 6.2. Architecture Trade-off Analysis 109
• 6.3. Comparative Analysis with Prior Research 111
• 6.3.1. Comparison of D-FLIB and Existing Centralized FL 111
• 6.3.2. Comparison between D-FLIB and Existing BFL 114
• 6.4. Future Research 124
• 6.4.1. Topic Chain Model 125
• 6.4.2. Future Research Direction 126
• 7. Conclusion 129
• References 131
• Abstract in Korean 138