Project onto probability simplex by default. Code from https://github.com/hsnamkoong/robustopt/blob/master/src/simple_projections.py See MIT

Run a single trial.

.Fix locations where we’ll put the patches.

Split data for meta update steps and other defenses.

Transfer data to autoML code. Lazy init due to additional libraries.

Construct according to the benchmark.

https://discuss.pytorch.org/t/cropping-a-minibatch-of-images-each-image-a-bit-differently/12247/5.

Get accuracy and loss for all targets on their intended class.

Create an LMDB database from the given pytorch dataset. https://github.com/Lyken17/Efficient-PyTorch/blob/master/tools/folder2lmdb.py Remove

Construct according to the triplet input key. Poisons are always the first n occurences of the given class. [This is the same setup a

Return only the pair (target, idx).

Choose targets from some label which will be poisoned toward some other chosen label.

Print info into console and into the stats object.

Construct according to random selection. The setup can be repeated from its key (which initializes the random generator). This method

https://discuss.pytorch.org/t/cropping-a-minibatch-of-images-each-image-a-bit-differently/12247/5.

Get accuracy of model relative to dataloader. Hint: The validation numbers in "base" and "target" explicitely reference the first label in intend

Save keys to .csv files.

Interface to connect to a kettle [data] child class.

Implement Main interface.

Implement Main interface.

Implement Chen et al. "Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering".

deepKNN as in Peri et al. "Deep k-NN Defense against Clean-label Data Poisoning Attacks". An overestimation factor of 2 is motivated as necessary

Implement the spectral signautres defense proposed by Tran et al. in "Spectral Signatures in Backdoor Attacks". https://proceedings.neurips.cc/pa

Enter.

Return model to CPU.

This is only called if all attributes of Subset are exhausted.

Call this only if all attributes of Subset are exhausted.

Call this only if all attributes of Subset are exhausted.

Getitem from https://pytorch.org/docs/stable/_modules/torchvision/datasets/cifar.html#CIFAR10. Args: index (int): Index

Getitem from https://pytorch.org/docs/stable/_modules/torchvision/datasets/cifar.html#CIFAR10. Args: index (int): Index

_getitem from https://pytorch.org/docs/stable/_modules/torchvision/datasets/mnist.html#MNIST. Args: index (int): Index R

_getitem from https://pytorch.org/docs/stable/_modules/torchvision/datasets/folder.html#DatasetFolder. Args: index (int): Index

Return a triplet of image, label, index.

Get from database. This is still unordered access. Cursor access would be a major hassle in conjunction with random shuffles.

Initialize with given specs...

Initialize with a given pytorch dataset.

Use as torchvision.datasets.ImageNet.

As torchvision.datasets.ImageNet except for additional keyword 'limit'.

Init with split, transform, target_transform. use --cached_dataset data is to be kept in memory.

Args: source and target size.

Initialize with a given pytorch dataset.

Initialize with given specs...

Cut-out with given alpha value. 0.66 is CIFAR-specific so that # s.t. sqrt(1 - 0.66) * 28 approx 16

Initialize with mixing factor alpha.

Initialize with a given data augmentation module.

Init with setup info.

Initialize empty victim.

Initialize with dict containing type and strength of attack and model info.

Init with width and num classes.

Initialize empty victim.

Initialize a model with given specs...

Init with network.

Return length via image paths.

Draw length from target dataset.

Step through a model epoch to in turn minimize target loss.

Take a step toward minmizing the current target loss.

Sanity check: Contructing data poisons by watermarking.

Implement the closure here.

Implement the closure here.

Implement the closure here.

Implement the closure here.

Implement the closure here.

Implement the closure here.

Validate a given poison by training the model and checking target accuracy.

Validate a given poison by training the model and checking target accuracy.

Validate a given poison by training the model and checking target accuracy.

Evaluate negative CrossEntropy for a gradient ascent.

Evaluate Gradient Alignment and descend.

Evaluate Gradient Alignment and descend.

Evaluate Metapoison.

Evaluate Gradient Alignment and descend.

Run a single trial.