Privacy, Utility and Robustness of Synthetic Tabular Data for Financial Machine Learning: A Multi-Seed Benchmark with Formal Privacy Analysis

Financial institutions want to share and develop models on synthetic tabular data, but published comparisons rarely test utility, fidelity, privacy, robustness and fairness together, with repeated runs and on real data. This study benchmarks six generators (a class-conditional Gaussian copula, kNN interpolation, a differentially private naive-Bayes synthesizer at two budgets, CTGAN and TVAE) on two public datasets: credit-card default (24,000 records) and card fraud (284,807 records, split chronologically). Each configuration was run with five seeds and evaluated by train-on-synthetic, test-on-real utility, fidelity, two membership-inference attacks, augmentation, perturbation robustness and subgroup error gaps, with Holm-adjusted paired tests and equivalence tests. Models trained on synthetic data lost 0.005–0.069 AUC against real-data models. kNN interpolation was closest to real-data utility but was the only generator that leaked membership (attack AUC 0.70 and 0.76), as a short proof of its distance bound anticipates; all others were at chance. CTGAN was worse than the copula (−0.049 and −0.017 AUC) and improved only with more training records and epochs, without reaching it. Synthetic augmentation never improved average precision. The private synthesizer plateaued for budgets of about 5 and above and degraded sharply below ε = 1. Code, data pointers and all per-run results are released.