ResNet generalization for CIFAR-like thingies. This is a minor modification of https://github.com/pytorch/vision/blob/master/torchvision/mode

Overwrite subset class to provide class methods of main class.

Hyperparameters used by this framework.

Cache a given dataset.

Implement LMDB caching and access. Based on https://github.com/pytorch/vision/blob/master/torchvision/datasets/lsun.py and https://github

Sanity check: do not actually attack - just use the random initialization.

Randomly patch pairs of classes as in Liam's implementation and optimize over these triggers.

Super-class CIFAR10 to return image ids with images.

Super-class CIFAR100 to return image ids with images.

Implement cutmix, a data augmentation combining cutout and mixup. https://arxiv.org/abs/1905.04899. This is fixed to nway=2 for now and hardcodes

GPU context for quick (code-wise) moves to and from GPU.

Overwrite torchvision ImageNet to change metafile location if metafile cannot be written due to some reason.

Overwrite torchvision ImageNet to limit it to less than 1mio examples. [limit/per class, due to automl restrictions].

Generate parameters for an experiment randomly. If --poisonkey is provided, then it will be used to seed the randomization.

Super-class MNIST to return image ids with images.

Crop the given batch of tensors at a random location. As discussed in https://discuss.pytorch.org/t/cropping-a-minibatch-of-images-each-image-a-b

Tiny ImageNet data set available from `http://cs231n.stanford.edu/tiny-imagenet-200.zip`. Author: Meng Lee, mnicnc404 Date: 2018/06/04 Re

Randomly patch pairs of classes as in Liam's implementation and optimize over these triggers. Allow the target patch to differ from the input pat

Implement a basic untargeted attack objective.

Implement limited steps for data poisoning via gradient alignment.

expand + depthwise + pointwise

ConvNetBN.

This is data augmentation via Cutout. https://arxiv.org/abs/1708.04552.

Implement limited steps for data poisoning via feature collision (with the bullseye polytope variant).

Implement limited steps for data poisoning via hidden trigger backdoor. Note that this attack modifies temp_targets as a side-effect!

Generate parameters for an experiment as specified in the data poisoning benchmark. https://github.com/aks2203/poisoning-benchmark

Generate parameters for an experiment based on a fixed triplet a-b-c given via --poisonkey. This construction replicates the experiment definitio

Implement limited steps for data poisoning via gradient alignment.

A meta-augmentation, returning the worst result from a range of augmentations. As in the orignal paper, https://arxiv.org/abs/2002.09024, thi

Trace a networks and then replace its module calls with functional calls. This allows for backpropagation w.r.t to weights for "normal" PyTorch n

Implement limited steps for data poisoning via MetaPoison.

This is data augmentation via mixup. https://arxiv.org/abs/1710.09412.

Randomly patch 2 classes.

Randomly patch 2 classes.

Randomly patch classes with patches consisting of randomly drawn image patches from the inputs.

Randomly patch pairs of images.

Randomly patch 2 classes.

Draw random patch shapes from the interval [0.5 * eps, 1.75 * eps]. This interval is slightly biased: Its mean length is not eps, but: eps: i

Sanity check: attack by watermarking.

Brew poison frogs variant with averaged feature matching instead of sums of feature matches. This is also known as BullsEye Polytope Attack.

Brew poison frogs variant with averaged feature matching instead of sums of feature matches. This is also known as BullsEye Polytope Attack.

Brew poison frogs poison with given arguments. “Double, double toil and trouble; Fire burn, and cauldron bubble.... Round about the caul

Brew passenger poison with given arguments. “Double, double toil and trouble; Fire burn, and cauldron bubble.... Round about the cauldro

Brew passenger poison with given arguments. Try to match the original image feature representation to hide the attack from filter defenses. T

Brew passenger poison with given arguments. Both the poison gradient and the target gradient are modified to be diff. private before calcuating t

Variant in which target gradients are matched separately.

Brew metapoison with given arguments. Note: This function does not work in single-model-multi-GPU mode, due to the weights being fixed to a singl

Reimplementation of metapoison using the "higher" library.

Reimplementation of metapoison using the "higher" library. This version also implements the "shared-batch" between target and inputs.

Brew poison with given arguments. “Double, double toil and trouble; Fire burn, and cauldron bubble.... Round about the cauldron go;

Brew poison with given arguments. “Double, double toil and trouble; Fire burn, and cauldron bubble.... Round about the cauldron go;

Implement model-specific code and behavior for multiple models on an unspecified number of GPUs and nodes. --> Running in concurrent mode!

Implement model-specific code and behavior for multiple models on a single GPU. --> Running in sequential mode!

Implement model-specific code and behavior for a single model on a single GPU. This is the simplest victim implementation.

Implement a variety of input-altering attacks.

Generate a dataset definition completely from file.

Brew poison with given arguments. Data class. Attributes: - trainloader - validloader - poisonloader - poison_ids - train

Implement model-specific code and behavior. Expose: Attributes: - model - optimizer - scheduler - criterion Methods

Brew poison with given arguments. Base class. This class implements _brew(), which is the main loop for iterative poisoning. New iterati