tapas.threat_models.attacker_knowledge

This file implements common examples of attacker knowledge of both the private dataset and the generator. This specifies elements (2.) and (3.) of threat models (see .base_classes.py).

Knowledge of the data is represented by AttackerKnowledgeOnData objects, which defines methods to sample training and testing datasets.

Knowledge of the generator is represented by AttackerKnowledgeOnGenerator objects, which is primarily a wrapper of generator.__call__ with a given number of synthetic records, but can be extended to more.

Classes

AttackerKnowledgeOnData()	Abstract base class that represents the knowledge that attackers have on real datasets, in the form of a prior from which they can sample training datasets for the attack.
AttackerKnowledgeOnGenerator()	Abstract base class that represents the knowledge that attachers have on the generator used to produce the synthetic datasets.
AttackerKnowledgeWithLabel()	Abstract base class that builds on AttackerKnowledgeOnData that adds the functionality of labeling the datasets.
AuxiliaryDataKnowledge([dataset, ...])	This attacker knowledge assumes access to some auxiliary dataset from which training datasets are sampled, as random subset of this auxiliary data.
BlackBoxKnowledge(generator, ...)	The attacker has access to the generator method with access to the generator has an exact black-box.
ExactDataKnowledge(training_dataset)	Also called worst-case attack, this assumes that the attacker knows the exact dataset used to generate
LabelInferenceThreatModel(...[, ...])	Label-inference Threat Model.
NoBoxKnowledge(generator, num_synthetic_records)	The attacker does not have access to the generator.
SilentIterator(*args, **kwargs)	SilentIterator implements the interface expected of iteration trackers, but does nothing.
UncertainBoxKnowledge(generator, ...[, ...])	The attacker has uncertain knowledge of the generator: they have access to the code, but not to some "parameters" of the code.