TECHNOLOGY PROFILE · TECH-0181
Opacus
Differentially private training library for PyTorch.
Reference link only — not reviewed this edition. This profile preserves the prepared catalog’s original descriptions. Advantages depend on the workload; comparisons are editorial, not benchmark results.
Where it has an edge
Per-sample gradients and privacy accounting reduce custom DP-SGD implementation work.
Best-fit problems
Research or deployment requiring a defined training privacy budget.
Tradeoffs & cautions
DP has utility and compute costs; correct accounting and threat assumptions are essential.
What to test before choosing
Write a threat model; test permissions/isolation and validate any privacy accounting with qualified review.
Runtime & hardware
Local, server or distributed
Access & licensing
Open-source cores; service/edition terms vary
Execution
Local / self-hosted
Lifecycle
Not independently audited
Comparison basis
Editorial problem-fit judgement; no head-to-head benchmark performed
Connected concepts
Alternatives to investigate
TensorFlow Privacy · Flower