Skip to main content

differential

Contains classes for marking differential privacy on models.

Classes​

DPModellerConfig​

class DPModellerConfig(    max_epsilon: float,    max_grad_norm: Union[float, List[float]] = 1.0,    alphas: List[float] = <factory>,    target_delta: float = 1e-06,    loss_reduction: "Literal[('mean', 'sum')]" = 'mean',    auto_fix: bool = True,):

Modeller configuration options for Differential Privacy.

Arguments

  • max_epsilon: The maximum epsilon value to use.
  • max_grad_norm: The maximum gradient norm to use. Defaults to 1.0.
  • alphas: The alphas to use. Defaults to floats from 1.1 to 63.0 (inclusive) with increments of 0.1 up to 11.0 followed by increments of 1.0 up to 63.0. Note that none of the alphas should be equal to 1.
  • target_delta: The target delta to use. Defaults to 1e-6.
  • loss_reduction: The loss reduction to use. Available options are "mean" and "sum". Defaults to "mean".
  • auto_fix: Whether to automatically fix the model if it is not DP-compliant. Currently, this just converts all BatchNorm layers to GroupNorm. Defaults to True.

Raises

  • ValueError: If loss_reduction is not one of "mean" or "sum".
info

max_epsilon and target_delta are also set by the Pods involved in the task and take precedence over the values supplied here.

Variables​

  • static alphas : List[float]
  • static auto_fix : bool
  • static fields_dict
  • static loss_reduction : Literal['mean', 'sum']
  • static max_epsilon : float
  • static max_grad_norm : Union[float, List[float]]
  • static nested_fields
  • static target_delta : float

DPPodConfig​

class DPPodConfig(    max_epsilon: float,    max_target_delta: float = 1e-06,    fields_dict: _StrAnyDict = <factory>,):

Pod configuration options for Differential Privacy.

Primarily used as caps and bounds for what options may be set by the modeller.

Arguments

  • max_epsilon: The maximum epsilon value to use.
  • max_target_delta: The maximum target delta to use. Defaults to 1e-6.

Ancestors​

  • bitfount.federated.privacy.differential._BaseDPConfig

Variables​

  • static fields_dict : Dict[str, Any]
  • static max_epsilon : float
  • static max_target_delta : float