Module: tf_privacy

TensorFlow Privacy library.

Modules

restart_query module: Implements DPQuery interface for restarting the states of another query.

tree_aggregation module: Tree aggregation algorithm.

v1 module: TensorFlow Privacy library v1 imports.

Classes

class DNNClassifier: DP version of tf.estimator.DNNClassifier.

class DPKerasAdagradOptimizer: Differentially private subclass of class tf.keras.optimizers.Adagrad.

class DPKerasAdamOptimizer: Differentially private subclass of class tf.keras.optimizers.Adam.

class DPKerasSGDOptimizer: Differentially private subclass of class tf.keras.optimizers.SGD.

class DPModel: DP subclass of tf.keras.Model.

class DPQuery: Interface for differentially private query mechanisms.

class DPSequential: DP subclass of tf.keras.Sequential.

class DiscreteGaussianSumQuery: Implements DPQuery for discrete Gaussian sum queries.

class DistributedDiscreteGaussianSumQuery: Implements DPQuery for discrete distributed Gaussian sum queries.

class DistributedSkellamSumQuery: Implements DPQuery interface for discrete distributed sum queries.

class GaussianSumQuery: Implements DPQuery interface for Gaussian sum queries.

class NestedQuery: Implements DPQuery interface for structured queries.

class NoPrivacyAverageQuery: Implements DPQuery interface for an average query with no privacy.

class NoPrivacyQuantileEstimatorQuery: Iterative process to estimate target quantile of a univariate distribution.

class NoPrivacySumQuery: Implements DPQuery interface for a sum query with no privacy.

class NormalizedQuery: DPQuery for queries with a DPQuery numerator and fixed denominator.

class QuantileAdaptiveClipSumQuery: DPQuery for Gaussian sum queries with adaptive clipping.

class QuantileEstimatorQuery: DPQuery to estimate target quantile of a univariate distribution.

class RestartQuery: DPQuery for SumAggregationDPQuery with a reset_state function.

class SumAggregationDPQuery: Base class for DPQueries that aggregate via sum.

class TreeCumulativeSumQuery: Returns private cumulative sums by clipping and adding correlated noise.

class TreeRangeSumQuery: Implements dp_query for accurate range queries using tree aggregation.

class TreeResidualSumQuery: Implements DPQuery for adding correlated noise through tree structure.

class VectorizedDPKerasAdagradOptimizer: Vectorized differentially private subclass of given class

class VectorizedDPKerasAdamOptimizer: Vectorized differentially private subclass of given class

class VectorizedDPKerasSGDOptimizer: Vectorized differentially private subclass of given class

Functions

compute_dp_sgd_privacy(...): Compute epsilon based on the given hyperparameters.

compute_rdp_single_tree(...): Computes RDP of the Tree Aggregation Protocol for a single tree.

compute_rdp_tree_restart(...): Computes RDP of the Tree Aggregation Protocol for Gaussian Mechanism.

compute_zcdp_single_tree(...): Computes zCDP of the Tree Aggregation Protocol for a single tree.

linearly_separable_labeled_examples(...): Generates num_examples labeled examples using separator given by weights.

logistic_dpsgd(...): Trains and validates private logistic regression model via DP-SGD.

logistic_objective_perturbation(...): Trains and validates differentially private logistic regression model.

make_dp_model_class(...): Given a subclass of tf.keras.Model, returns a DP-SGD version of it.

make_keras_optimizer_class(...): Given a subclass of tf.keras.optimizers.Optimizer, returns a DP-SGD subclass of it.

make_vectorized_keras_optimizer_class(...): Given a subclass of tf.keras.optimizers.Optimizer, returns a vectorized DP-SGD subclass of it.

single_layer_softmax_classifier(...): Trains a single layer neural network classifier with softmax activation.

synthetic_linearly_separable_data(...): Generates synthetic train and test data for logistic regression.

version '0.8.4'