Identical¶

Identical-gradient Byzantine attacks.

These attacks generate f_real references to the same newly created Byzantine gradient. The gradient is built from the average honest gradient plus a scaled attack direction.

Available attack directions are:

Bulyan: unit vector, optionally restricted to one coordinate.
Empire: negative average honest gradient.
Little: coordinate-wise standard deviation of honest gradients.

Use Case¶

Testing aggregation rules against attacks that submit identical malicious gradients from multiple Byzantine workers.

Properties¶

Identical gradients: All Byzantine workers submit the same gradient.
Direction-based: Attack direction is computed from honest gradients.
Factor optimization: Negative factors trigger automatic optimization.
Returns newly created tensors, does not alias honest input gradients.

param factor:: Attack scaling factor. Positive values use the provided factor directly. Negative integers trigger a line search over -factor evaluations. Defaults to -16.
type factor:: float or int, optional
param negative:: Whether to negate the selected factor. Defaults to False.
type negative:: bool, optional
param Example:
param ——-:
param >>> import torch:
param >>> from aggregators import average:
param >>> from attacks import little:
param >>> grad_honests = [torch.tensor([1.:
param 2.:
param 3.]):
param torch.tensor([4.:
param 5.:
param 6.])]:
param >>> byzantine_grads = little(:
param … grad_honests=grad_honests:

:param : :param … f_decl=1: :param : :param … f_real=1: :param : :param … defense=average: :param : :param … model=None: :param : :param … factor=1.5: :param : :param … ): :param >>> len(byzantine_grads): :param 1: