November 2019
Intermediate to advanced
346 pages
9h 36m
English
We begin Steps 1-3 by preparing and loading the MNIST dataset. Next, in Step 4, we import DPGradientDescentGaussianOptimizer, an optimizer that allows the model to become differentially private. A number of parameters are used at this stage, and these stand to be clarified. The l2_norm_clip parameter refers to the maximum norm of each gradient computed on an individual training datapoint from a minibatch. This parameter bounds the sensitivity of the optimizer to individual training points, thereby moving the model toward differential privacy. The noise_multiplier parameter controls the amount of random noise added to gradients. Generally, the more noise, the greater the privacy. Having finished this step, in Step 5, we define ...