Compute log(erfc(x)) with high accuracy for large x.

Precondition gradients with side information. Args: opt: Optimizer instance.

Setup the loss function for distillation. `crossentropy`, `L2` or `L1`. Args: name (str): type of the function R

Set the noise type. Args: noise_type (str): Type of noise.

Set the upper bound function. Args: backend (str): Backend for calculation. bound_type (str): Type of bound.

Assigns a function to self.distancefunc according to distancename Args: distancename: name of the function to calculat the distan

Assigns a class to self.optimizer_class according to optimiername Args: optimizername: name of optimizer, currently support `LBFG

Performs step 2 of the Hungarian algorithm. Determines whether it is possible to select one zero from each row and column. If

Performs step 3 of the Hungarian algorithm. Covers all zeros with the minimum number of horizontal or vertical lines. Args:

Performs step 4 of the Hungarian algorithm. Subtracts the minimum value from the elements not covered by the lines, and add t

Gets prediction and its membership label per each class from shadow models

Update side information for Adam optimizer. Args: opt: Optimizer instance.

Update side information for RMSprop optimizer. Args: opt: Optimizer instance.

* @brief Aggregates categorical column * * @param df * @param indices * @param column * @return std::string */

* @brief Aggregates continuous column * * @param df * @param indices * @param column * @return float */

Train the own local model to minimize the distance between the global logits and the output logits of the local model on the public dataset.

Wraps the given class in GANAttackClientWrapper. Returns: cls: a class wrapped in GANAttackClientWrapper

Wraps the given class in FreeRiderClientWrapper. Returns: cls: a class wrapped in FreeRiderClientWrapper

Wraps the given class in FoolsGoldServerWrapper. Returns: cls: a class wrapped in FoolsGoldServerWrapper

Attach the AdaDPS optimizer to the given class. Args: cls: Class to which AdaDPS optimizer will be attached. accountant: Pri

Wraps the given class in DistributedBackdoorAttackClientWrapper. Args: cls: Server class decomposed_trigger_rules ([function]): l

Wraps the given optimizer class in DPOptimizerWrapper. Args: accountant (BaseMomentAccountant): moment accountant l2_norm_clip (f

Attaches DPSGD (Differentially Private Stochastic Gradient Descent) functionality to the client class. Args: cls: Client class to wh

Wraps the given class in FoolsGoldServerWrapper. Returns: cls: a class wrapped in FoolsGoldServerWrapper

Wraps the given class in FreeRiderClientWrapper. Args: mu (float): mean of the gaussian distribution used to generate fake gradients

Wraps the given class in GANAttackClientWrapper. Args: target_label(int): index of target class generator (torch.nn.Module): Gene

Wraps the given class in GradientInversionServerWrapper. Args: x_shape: input shape of target_model. attack_function_on_receive (

Attaches a history attack to a client. Args: cls: The client class. lam (float): The lambda parameter for the attack. Return

Attaches a label flip attack to a client. Args: cls: The client class. victim_label: The label to be replaced. target_lab

Attaches a MAPF attack to a client. Args: cls: The client class. lam (float): The lambda parameter for the attack. base_m

Wraps the given class in ModelReplacementAttackClientWrapper. Args: cls: Client class Returns: cls: a class wrapped in ModelR

Attaches a normalization attack to a SplitNN model. Args: cls: The SplitNN model class. attack_criterion: The criterion for the a

Makes the client class communicate the encrypted gradients with paillier encryption scheme. Args: cls: client class pk: public ke

Attaches the Soteria wrapper to the client class. Args: cls: The client class to which Soteria will be attached. input_layer

Make the client class communicate the sparse gradients. Args: cls: client class k (int): strength of sparcity Returns:

Make the server class communicate the sparse gradients. Args: cls: server class Returns: cls: a class wrapped in SparseGradi

Execute the model inversion attack on the target model. Args: target_label (int): taregt label lam (float): step size

Computes the batch normalization regularizer loss. Args: feature_maps (list): List of feature maps. bn_layers (list): List of bat

* Bucket interface */

Calculate the upper bound of Renyi Differential Privacy (RDP) for one step. Args: alpha (float): Privacy parameter alpha

Compute log(A_alpha) for fractional alpha. 0 < q < 1.

Renyi Differential Privacy of the Sampled Gaussian Mechanism 3.3 Numerically Stable Computatio

* Modifiers */

* Modifiers */

Cross entropy loss for soft labels Based on https://discuss.pytorch.org/t/soft-cross-entropy-loss-tf-has-it-does-pytorch-have-it/69501/2 Args:

Distribute the logits of public dataset to each client.

Distribute the logits of public dataset to each client.

Distribute the logits of public dataset to each client.

Download the global logits from the server. Args: global_logit (torch.Tensor): the global logits from the server