TensorFlow Privacy: Learning with differential privacy for training data

Description

When evaluating ML models, it can be difficult to tell the difference between what the models learned to generalize from training and what the models have simply memorized. And that difference can be crucial in some ML tasks, such as when ML models are trained using sensitive data. Recently, new techniques have emerged for differentially private training of ML models, including deep neural networks (DNNs), that used modified stochastic gradient descent to provide strong privacy guarantees for training data.

Those techniques are now available, and they’re both practical and can be easy to use. This said, they come with their own set of hyperparameters that need to be tuned, and they necessarily make learning less sensitive to outlier data in ways that are likely to slightly reduce utility. Úlfar Erlingsson explores the basics of ML privacy, introduces differential privacy and why it’s considered a gold standard, explains the concrete use of ML privacy and the principled techniques behind it, and dives into intended and unintended memorization and how it differs from generalization.