Federated Learning with Differential Privacy for Credit Risk Assessment in the Moroccan Banking Sector: A Data-Driven Approach for Secure Open Banking

The rapid expansion of Open Banking in Morocco, accelerated by Bank Al-Maghrib regulatory frameworks and the digital transformation of financial services, creates systemic challenges for credit risk assessment across distributed banking networks. This paper proposes a federated learning (FL) architecture reinforced by differential privacy (DP) mechanisms for collaborative credit risk modelling in the Moroccan banking sector. The framework enables financial institutions to jointly train predictive models on distributed customer data without compromising individual privacy or violating data sovereignty constraints. Drawing on empirical data from a survey of 500 clients and 25 expert interviews conducted in the Rabat-Salé-Kénitra (RSK) region, we design a privacy-preserving gradient aggregation protocol adapted to the heterogeneous structures of Moroccan retail banking portfolios. Our federated model (FL-DP-AC) achieves an AUC-ROC of 0.847, representing a 14.3% improvement over centralised baselines, while maintaining a privacy budget (ε) of 1.2 under the Gaussian mechanism, meeting privacy thresholds imposed by Moroccan data protection law (Law No. 09-08). Results validate the feasibility of federated credit scoring as a secure and interoperable foundation for Open Banking ecosystems, and contribute a blueprint for AI-driven financial inclusion across the African banking sector.