tapas.threat_models.base_classes

Threat Models describe the assumptions under which an attack takes place.

Threat Models are composed of three elements:

  1. What the attacker aims to infer (e.g., membership, attribute).

  2. What the attacker knows about the generator (no-, black-, white-box).

  3. What the attacker knows about the training dataset.

A threat model thus describes what an attack wants to predict, and how it can do that. It also describes how to _evaluate_ the success of an attack. For instance, for a black-box membership inference attack with auxiliary data, the attacker is able to run the generator on datasets samples from the auxiliary data, which may or may not contain a target record. The evaluation of the attack is performed on datasets from a disjoint dataset (test set), from which training datasets are sampled, with or without the target record.

Classes

ThreatModel()

Abstract base class for a threat model.

TrainableThreatModel()

Some threat models additionally define a way to train attacks with synthetic datasets generated using the attacker's knowledge.