DP-SGD — Differentially Private SGD

Why it matters

DP-SGD is the gold standard for private ML. Understanding shapes compliant training.

The architecture

Per-example gradient: compute for each sample separately.

Clip: cap gradient L2 norm to C.

Aggregate + noise: sum, add Gaussian noise scaled by C.

How it works end to end

Privacy accountant: track cumulative ε across training.

Hyperparameters: clipping norm C, noise multiplier σ, batch size.

Cost: slower training (per-example gradients); quality drop.

Libraries: Opacus (PyTorch), TF Privacy.