Privacy-preserving federated credit risk models: evaluating differential privacy and homomorphic encryption techniques

The adoption of machine learning in the financial sector requires solutions that ensure secure data collaboration while maintaining regulatory compliance. Federated Learning (FL) offers a decentralized alternative to centralized training; however, it remains vulnerable to information leakage through gradient sharing. This study proposes a privacy-preserving FL framework for loan approval prediction and evaluates three privacy configurations: Standard FL, Differential Privacy-enabled FL (DP-FL), and Homomorphic Encryption-enabled FL (HE-FL). A real-world loan dataset distributed across five simulated clients was used to assess model performance, computational overhead, and privacy guarantees. Differential Privacy was implemented using the Gaussian mechanism, with formally computed cumulative privacy budgets (ε = 14.13, 8.65, and 5.74) derived using Rényi Differential Privacy accounting over 1000 local epochs and five communication rounds. Experimental results show that Standard FL achieved the highest accuracy (≈ 91%) but provided no confidentiality guarantees. HE-FL preserved accuracy (≈ 90%) while ensuring encrypted computation at the cost of increased overhead. DP-FL demonstrated predictable privacy-utility behaviour, where ε = 8.65 yielded the most balanced performance (≈ 87-90%) with enforceable privacy guarantees and moderate computational cost. The findings confirm that privacy-preserving FL is feasible for regulated financial applications and that privacy performance can be tuned based on operational and regulatory requirements.