Assessment of data privatization methods under local differential privacy in the banking sector
Laura Valentina López Ardila · instname:Universidad de los Andes · 2025
In today’s data-driven world, protecting personal information has become a critical challenge as organizations increasingly rely on large-scale data collection for decision-making. Local Differential Privacy (LDP) offers a compelling solution by allowing users to independently perturb their data before sharing it, ensuring strong privacy guarantees without relying on a trusted third party. This study evaluates the impact of LDP mechanisms on predictive performance and structural preservation in a financial customer churn model. Multiple approaches—Direct Encoding (DE), Optimized Unary Encoding (OUE), Optimized Local Hashing (OLH), Treshold Histogram Encoding (THE), Two-Dimensional Grid (TDG), Hybrid-Dimensional Grid(HDG) and Local Differential Privacy Principal Component Analysis (LDP-PPA)—were compared under different privacy budgets and perturbation settings. Results show that model utility behaves nonlinearly as the most influential variables are privatized, emphasizing the value of adaptive privacy strategies. Aggregated mechanisms such as OLH and THE consistently achieved better privacy–utility trade-offs, while LDP–PPA effectively preserved feature correlations through dimensionality reduction. Evaluation through norms of the correlation differences confirmed that aggregated or hash-based methods cause less distortion than direct randomization. These findings demonstrate that LDP can be integrated into financial analytics pipelines to protect user data without compromising utility.