DP-SGD mechanisms¶
DP-SGD adds calibrated noise to a clipped gradient at every step.
Privacy is composed step-by-step: each call to the noise mechanism is
a separate DpProcess, multiplied by the number of training steps in
the accountant.
Mechanisms¶
- Gaussian — independent Gaussian noise on the clipped gradient; the canonical DP-SGD release.
Pairing with sampling¶
DP-SGD typically pairs the Gaussian mechanism with Poisson subsampling
(opake.dpsgd.sampling.PoissonSampler). The accounting form is
opake.dpsgd.accounting.poisson(opake.dpsgd.accounting.gaussian(nm),
sample_rate=q); multiply by * num_steps for full-training privacy.
See also¶
- DP-SGD end-to-end guide — full training pipeline.
- DP-FTRL mechanisms — the other family of noise mechanisms in Opake (correlated, whole-process).