Scalable DP-SGD: Shuffling vs. Poisson Subsampling

A common approach for private training of differentiable models, such as neural networks, is to apply first-order methods with noisy gradients.