Evaluating Privacy–Utility Trade‑Offs on Small Clinical Datasets with PyTorch and Opacus
Updated on January 29, 2026 15 minutes read
Frequently Asked Questions
How much clinical expertise do I need to use DP‑SGD effectively?
You can implement DP‑SGD with strong ML skills, but you need domain input to choose labels, metrics, and acceptable error rates. In healthcare, the “right” operating point is usually determined by workflow constraints and risk tolerance, not by ML convention.
Can DP‑SGD work when I only have a few hundred patient records?
It can, but the trade‑offs are sharper. You should expect higher variance, faster privacy spending, and a greater need for conservative models, careful validation, and uncertainty reporting.
Should I tune `max_grad_norm` a lot, or mostly tune `noise_multiplier`?
Many teams start with max_grad_norm near 1.0 and tune noise and batch size first, because clipping thresholds often become stable after early training. But when clipping clearly harms convergence, experimenting with a few clipping values is worthwhile.
Is it safe to log per-sample gradient norms for debugging?
Not in a private setting. Opacus documentation explicitly notes that per-sample gradient norms are not privatized and should only be used for debugging or non-private contexts.