BCFL-DAI-PUB: Blockchain-integrated federated learning with dynamic adaptive incentives for privacy-utility balanced credit risk modelling
- 1Vellore Institute of Technology, School of Computer Science and Engineering, Vellore, Tamil Nadu, India.
- 2Vellore Institute of Technology, School of Computer Science and Engineering, Chennai, Tamil Nadu, India.
Abstract
Accurate prediction of loan defaulters through collaborative analysis of financial data across multiple institutions is critical for maintaining financial stability and strengthening credit risk management. Although Federated Learning (FL) enables organizations to train models without exchanging raw data, its reliance on a centralized server introduces serious trust concerns, including vulnerability to gradient inversion and membership inference attacks. Blockchain-based Federated Learning (BCFL) addresses these challenges by cryptographically protecting and aggregating model updates through a smart contract, thereby providing a tamper-resistant learning process. In spite of these security enhancements, current BCFL frameworks do not ensure sustained client participation in real-world settings. Moreover, while traditional incentive mechanisms promote truthful participation and discourage free-riding and malicious behavior, they mainly reward model utility and neglect privacy contribution. To mitigate this limitation, we propose BCFL-DAI-PUB, a blockchain-federated learning framework integrating an innovative Dynamic Adaptive Incentive (DAI) mechanism to achieve a Privacy–Utility Balance (PUB). The proposed DAI dynamically applies three complementary incentive components, such as SCAFFOLD-based correction, adaptive differential privacy (DP-Boost), and monetary rewards, to jointly optimize learning utility and privacy protection. This work promotes honest participation, discourages dishonest behavior, and enables rapid recovery from malicious attacks, thereby improving global convergence and overall model effectiveness. The proposed BCFL-DAI-PUB achieves a final AUC of 0.91, improving upon the HE-based BCFL-PS-HER baseline (0.86) by 1.06×. By replacing homomorphic encryption with adaptive differential privacy and combining it with a dynamic incentive mechanism, the proposed framework improves both efficiency and resilience over the baseline, which relies on static reputation-based coordination. Compared with the baseline BCFL-PS-HER model, BCFL-DAI-PUB recovers from malicious disruption 1.82× faster and achieves approximately 1.22× higher MIA resistance, while the baseline incurs approximately 1.09× higher communication overhead and 1.14× higher storage overhead. These results show that BCFL-DAI-PUB improves predictive performance, strengthens privacy protection, reduces system cost, and provides more robust operation in blockchain-enabled federated credit-risk modelling.
Related
- Advancing digital agri-food systems through integrated AI, robotics, IoT, and blockchain with a bibliometric analysis
- Increased monetary incentives enhance goal-directed reinforcement learning but fail to suppress automatic learning processes
- Data Amount Heterogeneity Improves Performance in Federated Learning for Neuroimaging
- EduBot: Integrating Federated Learning and RAG for Educational Document Analysis
- Serum proteins predict incident inflammatory bowel disease 14 years before diagnosis: a machine-learning and Cox modelling in the UK Biobank
BibTeX
@article{ramachandra2026bcfldaipub,
title = {{BCFL-DAI-PUB: Blockchain-integrated federated learning with dynamic adaptive incentives for privacy-utility balanced credit risk modelling}},
author = {Padmaja Ramachandra and Santhi Vaithiyanathan},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0361223},
url = {https://doi.org/10.1371/journal.pone.0361223},
note = {CC BY},
}
RIS
TY - JOUR
TI - BCFL-DAI-PUB: Blockchain-integrated federated learning with dynamic adaptive incentives for privacy-utility balanced credit risk modelling
AU - Padmaja Ramachandra
AU - Santhi Vaithiyanathan
PY - 2026
DA - 2026/10/29
JO - PLOS One
DO - 10.1371/journal.pone.0361223
UR - https://doi.org/10.1371/journal.pone.0361223
AB - Accurate prediction of loan defaulters through collaborative analysis of financial data across multiple institutions is critical for maintaining financial stability and strengthening credit risk management. Although Federated Learning (FL) enables organizations to train models without exchanging raw data, its reliance on a centralized server introduces serious trust concerns, including vulnerability to gradient inversion and membership inference attacks. Blockchain-based Federated Learning (BCFL) addresses these challenges by cryptographically protecting and aggregating model updates through a smart contract, thereby providing a tamper-resistant learning process. In spite of these security enhancements, current BCFL frameworks do not ensure sustained client participation in real-world settings. Moreover, while traditional incentive mechanisms promote truthful participation and discourage free-riding and malicious behavior, they mainly reward model utility and neglect privacy contribution. To mitigate this limitation, we propose BCFL-DAI-PUB, a blockchain-federated learning framework integrating an innovative Dynamic Adaptive Incentive (DAI) mechanism to achieve a Privacy–Utility Balance (PUB). The proposed DAI dynamically applies three complementary incentive components, such as SCAFFOLD-based correction, adaptive differential privacy (DP-Boost), and monetary rewards, to jointly optimize learning utility and privacy protection. This work promotes honest participation, discourages dishonest behavior, and enables rapid recovery from malicious attacks, thereby improving global convergence and overall model effectiveness. The proposed BCFL-DAI-PUB achieves a final AUC of 0.91, improving upon the HE-based BCFL-PS-HER baseline (0.86) by 1.06×. By replacing homomorphic encryption with adaptive differential privacy and combining it with a dynamic incentive mechanism, the proposed framework improves both efficiency and resilience over the baseline, which relies on static reputation-based coordination. Compared with the baseline BCFL-PS-HER model, BCFL-DAI-PUB recovers from malicious disruption 1.82× faster and achieves approximately 1.22× higher MIA resistance, while the baseline incurs approximately 1.09× higher communication overhead and 1.14× higher storage overhead. These results show that BCFL-DAI-PUB improves predictive performance, strengthens privacy protection, reduces system cost, and provides more robust operation in blockchain-enabled federated credit-risk modelling.
SN - 1932-6203
ER -