Generative AI for Credit Risk Management: Risk-Conditional Diffusion with Holdout-Calibrated Release Gates for Empirically Privacy-Audited Synthetic Credit Data

Generative AI is increasingly used to share and augment confidential credit data, but high-fidelity generators can reveal which borrowers they were trained on. We study this tension on three public credit risk datasets and propose PG-Diff (proximity-gated diffusion), which couples a risk-conditional tabular diffusion model, combining mass-point encoding of zero-inflated amounts with classifier-free guidance on the default label, with a generator-agnostic release gate. The gate enforces absolute proximity quotas and a relative proximity balance, both calibrated on held-out real records so that records drawn independently of the training data pass them in expectation. Adopting the fidelity, utility and privacy evaluation framework of Yang and Yang, extended with low-false-positive membership inference and a model-risk agreement metric, we compare SMOTE, Gaussian copula, TVAE, CTGAN and diffusion releases over up to three random splits. Conditional diffusion reaches train-on-synthetic AUCs of 0.751, 0.845 and 0.748 on Taiwan, Give Me Some Credit and German (real-data models: 0.762, 0.834, 0.778), gives the most realistic records among the deep generators, and shows membership-inference AUCs close to 0.5 on the two larger datasets. Interpolation-based SMOTE attains similar utility but leaks heavily (attack AUC up to 0.76, true-positive rate at 1% false positives up to 0.58); the gate lowers that rate by 43% on average and flags SMOTE through relaxation rates of up to 97%. For non-interpolating generators the gate costs at most 0.039 AUC, and in a memorization stress test it moves the DCR share from 0.595 to 0.472. Holdout-calibrated release gates thus provide a generator-agnostic empirical safeguard that complements formal mechanisms such as differential privacy.