Gaussian DP for Reporting Differential Privacy Guarantees in Machine Learning

This position paper advocates for the adoption of non-asymptotic Gaussian Differential Privacy (GDP) as a more accurate method for reporting differential privacy guarantees in machine learning algorithms like DP-SGD. It introduces open-source numerical accountants that can compute privacy profiles and $f$-DP curves with high precision, demonstrating that GDP can effectively capture the privacy profiles of DP-SGD and other algorithms with minimal error. This approach is significant for practitioners as it provides clearer and more reliable privacy guarantees, thereby enhancing the robustness of privacy-preserving machine learning systems.