Federated Learning for Shariʿah-Compliant Credit Scoring in Islamic Banks: Preserving Data Privacy and Eliminating Riba-Related Features

Maisuna Mustapha Yahya (Department of Islamic Studies, Faculty of Arts, University of Abuja)
Akanbi Bello Muhammed (Department of Computer Science, Faculty of Science, University of Abuja)
Alwy Ahmed Mohamed (Department of Islamic Education, Universitas Muhammadiyah Surakarta, Indonesia, and The Agakhan High School, Mombasa)

Demak Universal Journal of Islam and Sharia
Vol. 4 No. 01 (2026): Demak Universal Journal of Islam and Sharia

Objective: The rapid expansion of Islamic banking across Africa, Southeast Asia, and the GCC region requires credit assessment frameworks that are privacy-preserving, data-efficient, and strictly Shariʿah-compliant. This study proposes FedIslami, a novel federated learning framework designed to train credit-scoring models collaboratively across distributed Islamic bank branches without transmitting sensitive raw client records. Theoretical framework: The study is theoretically grounded in the Islamic principle of Amānah (trusteeship of client data) and Maqāsid al-Shariʿah objectives, establishing an ethical foundation for data privacy, algorithmic justice, and religious compliance within modern financial engineering. Literature Review: Existing literature highlights that conventional machine learning models violate data trusteeship through centralised data aggregation and systematically incorporate interest-rate exposure metrics, debt-service-coverage ratios, and other Ribā-indexed features that are impermissible under Shariʿah law. Method: FedIslami systematically excises fourteen identified Ribā-contaminated features, replacing them with Shariʿah-permissible analogues. Utilising FedAvg with differential privacy (DP-FedAvg) and homomorphic encryption secure aggregation, the model was validated on 214,736 anonymised financing records (2019–2024) across nine Islamic banks in Nigeria, Malaysia, UAE, and Bangladesh. Results: FedIslami achieved outstanding predictive accuracy with an AUC of 0.894, a Gini coefficient of 0.788, and a Kolmogorov-Smirnov statistic of 0.581, significantly outperforming non-federated centralised baselines (0.871) and conventional federated models retaining Ribā features (0.842). The audit confirmed full compliance with AAOIFI Governance Standard No. 7 and IFSB-16. Implications: This research demonstrates that privacy-preserving federated learning effectively bridges performance gaps between modern Islamic and conventional credit-scoring systems while simultaneously reinforcing institutional ethical and religious data-governance obligations. Novelty: FedIslami introduces the first dual-layer privacy-preserving federated learning architecture that simultaneously addresses client privacy and algorithmic Shariʿah compliance through systemic Ribā feature elimination and Maqāsid-aligned variable substitution.