tf_privacy.logistic_dpsgd

Trains and validates private logistic regression model via DP-SGD.

tf_privacy.logistic_dpsgd(
    train_dataset: datasets.RegressionDataset,
    test_dataset: datasets.RegressionDataset,
    epsilon: float,
    delta: float,
    epochs: int,
    num_classes: int,
    batch_size: int,
    num_microbatches: int,
    clipping_norm: float
) -> List[float]

The training is based on the differentially private stochasstic gradient descent algorithm implemented in TensorFlow Privacy.

Args
train_dataset	consists of num_train many labeled examples, where the labels are in {0,1,...,num_classes-1}.
test_dataset	consists of num_test many labeled examples, where the labels are in {0,1,...,num_classes-1}.
epsilon	epsilon parameter in (epsilon, delta)-DP.
delta	delta parameter in (epsilon, delta)-DP.
epochs	number of training epochs.
num_classes	number of classes.
batch_size	the number of examples in each batch of gradient descent.
num_microbatches	the number of microbatches in gradient descent.
clipping_norm	the gradients will be normalized by DPKerasAdamOptimizer to have l2-norm at most clipping_norm.

Returns
List of test accuracies (one for each epoch) on test_dataset of model trained on train_dataset.