Differential Privacy for LLM Training

Why it matters

DP formalizes privacy in ML training. Understanding shapes privacy-sensitive training.

The architecture

Clip per-example gradients.

Add Gaussian noise.

Track privacy budget epsilon.

How it works end to end

DP-SGD: standard method.

Epsilon: privacy budget (smaller = more private).

Larger models + data help.

Google + Apple use DP for on-device.