Input Optimization

While other implementations (such as advertorch) exist, this one is meant to be as easy, accessible, informative and modular as training model. In fact, this implementation uses fastai’s Learner class, and inherits its functionality, such as the progress bar, the losses table, and even early stopping and lr scheduling. Useful for evasion and feature-collision attacks.

---

source

InputOptimizer

 InputOptimizer (model:fastai.torch_core.Module,
                 pert_cb:Type[__main__.PerturbationCallback],
                 loss:Callable=FlattenedLoss of CrossEntropyLoss(),
                 lr:float=None, targeted:bool=False, min_delta:float=0.01,
                 min_lr:float=1e-06, epoch_size:int=10, n_epochs:int=4)

Constructs adversarial examples: slightly perturbed inputs that fool classification models

	Type	Default	Details
model	Module
pert_cb	typing.Type[main.PerturbationCallback]
loss	typing.Callable	FlattenedLoss of CrossEntropyLoss()
lr	float	None	pass None to try pick lr based on other parameters
targeted	bool	False	Whether the constructed inputs should be classified as the specified targets or not
min_delta	float	0.01	Minimum loss delta for ReduceLROnPlateau and EarlyStoppingCallback
min_lr	float	1e-06	Minimum lr for ReduceLROnPlateau
epoch_size	int	10	Affects how often epoch-callbacks are called (e.g. Recorder`` andEarlyStoppingCallback`)
n_epochs	int	4

---

source

PerturbationCallback

 PerturbationCallback (after_create=None, before_fit=None,
                       before_epoch=None, before_train=None,
                       before_batch=None, after_pred=None,
                       after_loss=None, before_backward=None,
                       after_cancel_backward=None, after_backward=None,
                       before_step=None, after_cancel_step=None,
                       after_step=None, after_cancel_batch=None,
                       after_batch=None, after_cancel_train=None,
                       after_train=None, before_validate=None,
                       after_cancel_validate=None, after_validate=None,
                       after_cancel_epoch=None, after_epoch=None,
                       after_cancel_fit=None, after_fit=None)

Manages the input perturbation for an InputOptimizer

---

source

InputOptimizer.perturb

 InputOptimizer.perturb (dsets)