Differential privacy strategies for data analytics in the banking sector

Financial institutions face growing pressure to protect customer data while maintaining the ability to obtain business insights through data analysis. To tackle this challenge, Differential Privacy has risen in recent years as a promising solution to guarantee user data privacy. However, the utility of the analytical models trained employing Differential Privacy remains an open question. In this paper, we consider two main approaches: the private-model workflow, employing Centralized Differential Privacy; and the private-data workflow, implementing Local Differential Privacy. We test these two alternatives on an open bank marketing dataset, training classification models and comparing them against a non-private baseline. Findings show a sharp trade-off between privacy and utility in the private-data workflow, which allow for usable models under moderate privacy guarantees (privacy budget in the range 3-5). On the other hand, in the private-model case, adjusting training configurations can help reduce the impact of privacy constraints. Experiments show the sample size, batch size and learning rate to be the most influential hyperparameters, while the noise multiplier and the gradient clipping norm (related to the privacy mechanism) show limited impact. The results provide practical guidance to apply differential privacy in financial environments, supporting both regulatory compliance and effective data use.