Gradient perturbation: For a parametric function fθ(x) parameterized by θ and loss function L(fθ(x),y), usual mini-batched first-order optimizers update θ using gradients gt = 1 N