Privacy Amplification by Sampling
You'll learn: How randomly selecting only a subset of your training data for each update can significantly strengthen differential privacy guarantees without changing your algorithm.
The Core Idea
When training with differentially private SGD (DP-SGD), you don't need to use all your data in every iteration. By randomly sampling a small batch, you reduce the chance that any individual's data influences the model—and this improves your privacy guarantee!
Think of it like this: if a spy is trying to figure out whether Alice's data was in the training set, but each gradient update only uses 1% of the data randomly, the spy has to work much harder. Most updates won't touch Alice's data at all, making her presence harder to detect.
Poisson Sampling Analysis
The mathematical foundation relies on Poisson sampling: instead of fixed-size batches, imagine each data point is independently included with probability q (the sampling rate). This randomness "amplifies" privacy.
Key result: If your algorithm provides ε-differential privacy when run on the full dataset, sampling with probability q reduces the effective privacy loss to approximately q·ε (for small q).
This is huge! If you only sample 1% of data (q = 0.01), your privacy cost per iteration drops by roughly 100×. Over many training steps, this compounds through composition theorems, letting you train longer or achieve better privacy with the same noise.
In practice, DP-SGD implementations automatically account for this amplification when computing your total privacy budget.
Key Takeaway: Random subsampling isn't just computationally efficient—it's a powerful privacy amplifier that makes each gradient update's privacy cost proportional to the sampling rate, dramatically improving your overall privacy guarantee.