Adversary

class olympus.metrics.adversary.ClassifierAdversary(frequency_new_epoch: int = 0, frequency_new_batch: int = 0, frequency_new_trial: int = 0, priority: int = 0, epsilon: float = 0.25, accuracies: list = <factory>, losses: list = <factory>, distortions: list = <factory>, loader: torch.utils.data.dataloader.DataLoader = None) → None

Bases: olympus.observers.observer.Observer

Simple Adversary Generator from arxiv Measure how robust a network is from adversary attacks

\[adversary(image) = image + epsilon * sign(grad(cost(theta, image, t), image)\]

An adversary takes as input an image and returns a modified image that will try to induce an error on the classifier.

References

[1]	Ian J. Goodfellow, Jonathon Shlens, Christian Szegedy. “Explaining and Harnessing Adversarial Examples”, 20 Dec 2014

Attributes:

epislon: float = 0.25 (for mnist) 0.07 (for ImageNet)

Epsilon corresponds to the magnitude of the smallest bit of an image encoding converted to real number

Methods

every(*args[, epoch, batch])	Define how often this metric should be called
load_state_dict(state_dict)	Load a state dictionary to resume a previous training
on_end_train(task[, step])	Called at the end of training after the last epoch
on_new_batch(task, step[, input, context])	Called after a batch has been processed
on_new_epoch(task, epoch, context)	Called at the end of an epoch, before a new epoch starts
on_new_trial(task, step, parameters, uid)	Called after a trial has been processed
on_start_train(task[, step])	Called on ce the training starts
state_dict()	Return a state dictionary used to checkpointing and resuming
value()	Return the key values that metrics computes

adversarial
on_end_batch
on_end_epoch

accumulator = 0

adversarial(task, batch, target)

count = 0

distortion = 0

epsilon = 0.25

load_state_dict(state_dict)

Load a state dictionary to resume a previous training

loader = None

loss = 0

on_end_batch(task, step, input, context)

on_end_epoch(task, epoch, context)

on_end_train(task, step=None)

Called at the end of training after the last epoch

state_dict()

Return a state dictionary used to checkpointing and resuming

time = <olympus.utils.stat.StatStream object>

value()

Return the key values that metrics computes