Why it matters
DP-SGD is workhorse of private training. Understanding shapes implementation.
Advertisement
The architecture
Per-example gradient clipping.
Add Gaussian noise scaled by clip.
Poisson subsampling per batch.
Advertisement
How it works end to end
Clip: L2 norm to C.
Noise: Gaussian with std = sigma * C.
Poisson sampling: each ex included with prob q.
Moments accountant: track epsilon.