ResNet block.

State of a model, including optional epoch and optimizer.

Simple wrapper for time.clock().

Attack a model.

Start Tensorboard for an attack run.

Simple singleton log implementation with different drivers.

Train a model.

Start Tensorboard for a specific training run.

Dataset consisting of adversarial examples.

Evaluation on adversarial and clean examples.

Adversarial training.

Configuration for adversarial training.

Interface for adversarial training.

Generic attack.

Configuration for attacks.

Regular attack interface.

Corner Attack from https://github.com/fra31/sparse-imperceivable-attacks/blob/master/cornersearch_attacks.py.

Random sampling.

Random sampling.

Adversarial frames.

Random sampling.

Implementation of untargetetd PGD attack.

Implementation of untargetetd PGD attack.

Random sampling.

Reference implementation of the PGD attack as originally used by Madry et al.; the implementation is based on the following TensorFlow code:

Random sampling.

Implementation of untargetetd PGD attack.

Projection onto a box defined by minimum and maximum.

Wrapper for clamp.

Base classifier.

General, clean dataset used for training, testing and attacking.

Evaluation on clean examples.

Clip augmenter.

Confidence-calibrated adversarial training.

Configuration for confidence calibrated training.

Interface for adversarial training.

Out-of-distribution evaluation.

Evaluation on adversarial and clean examples.

Entropy computation based on logits.

Fixed initialization.

Fixed LeNet architecture, working on MNIST architectures only.

Flatten module.

Convert float32 to uint8; used for some augmenters that only work on uint8 or float32.

Generates a frame location.

Initialization using random noise; does not enforce any constraints, projections should be used instead.

Gaussian convolution. See https://pytorch.org/docs/stable/nn.html.

Abstract initialization.

L_0 norm.

L_0 projection.

Uniform initialization on L_1 sphere.

L_1 norm.

L_1 projection.

Uniform initialization, wrt. norm and direction, in L_1 ball.

Uniform initialization, wrt. volume, in L_1 ball.

L_2 norm.

L_2 projection.

Uniform initialization, wrt. norm and direction, in L_2 ball.

Uniform initialization, wrt. volume, in L_2 ball.

L_inf norm.

L_inf projection.

Standard L_inf initialization as by Madry et al.

Uniform initialization, wrt. norm and direction, in L_inf ball.

Uniform initialization, wrt. volume, in L_inf ball.

General LeNet classifier.

Defines log level.

Wrap a simple message.

MLP classifier.

Generates a frame location.

Abstract norm.

Normal training.

Configuration for normal training.

Interface for normal training for expeirments.

Normalization layer to be learned.

Abstract objective.

Pre-Activation ResNet block.

Pre-Activation ResNet bottleneck block.

More or less fixed pre-activation ResNet.

Abstract projection.

Combination of multiple initializers.

Simple classifier.