The ε Dilemma Balancing Privacy and Utility in Differential Privacy

This paper investigates the trade- off between privacy preservation and model utility in machine learning under differential privacy mechanisms. It compares two widely used approaches, the Laplace and Gaussian mechanisms, across six real -world classificatio n datasets and four representative models. The analysis examines how different privacy configurations affect both predictive performance and data protection in practical settings. Utility is measured using accuracy, precision, recall, and F1 score, while privacy protection is evaluated with distance- and risk- based indicators that quantify data exposure. The results show that the Laplace mechanism generally achieves higher utility at moderate privacy budgets (ε ≈ 6 –15), particularly in high- dimensional datasets, while maintaining acceptable privacy levels. In contrast, the Gaussian mechanism provides stronger privacy guarantees at lower budgets but causes greater performance degradation, especially for sparse or small datasets. Overall, this study offers pra ctical guidance for selecting and tuning privacy mechanisms based on dataset characteristics such as dimensionality, sparsity, and sample size. The proposed empirical framework can serve as a reference for applying differential privacy in real-world AI applications.