Quick-Start Guide

This guide is intended to get you familiar with the library and get things up and running as quickly as possible. Here we only describe how individual attacks with a single attack configuration are executed. For information about the attack runner and structured pentesting, see attack runner.

Library Structure

The library is kept very simple. You can execute an attack already with two function calls. And to generate an attack report, you need just one more. For example, for the privacy.mia attack, it would look as follows (for details about the function calls, see the documentation of the individual attacks):

1. Create an attack object: attack_object = mia.Mia(...)
2. Run the attack: attack_object.run(...)
3. Create an attack report (optional): attack_object.create_attack_report(...)

That way, also multiple instances of a model can be attacked to ensure that it is not the stochastic learning process that leads to a privacy-preserving and/or robust model but the chosen hyperparameters. For this purpose, you must pass a list of target model instances to the constructor of the attack object. To generate balanced training datasets for such a set of target models, you can use pepr.utilities.assign_record_ids_to_target_models().

Note

In the current version, the report generator can only generate reports for attacks on individual target models. If the report generator is called on an attack object with multiple target models, it generates an attack report for the first target model.

Note

To call attack_object.create_attack_report(pdf=True) you need a full LaTeX installation on your system.

Logging

For logging, we use Python’s standard library. For an example of how logging can look like, check out the example notebooks. For more details, please refer to the howto of the standard library.

Rapidly Re-Executing Attacks

We want to point out that you can re-execute some attacks without recomputing all the steps of the attack. For example, the privacy.gmia attack can be re-executed without having to re-train the expensive reference models. This can be done by calling the attack_object.run() method again with the load_pars parameter. For details, please consult the documentation of the corresponding attack.

Warning

Please use this feature only if you know what you are doing. Otherwise, you can easily get unexpected results.

Examples

For full and concrete examples checkout our example notebooks:

Notebook	Description	Google Colab
mia_tutorial	Attack a single target model with mia.
direct_gmia_tutorial	Attack a single target model with direct-gmia.
foolbox_tutorial	Run attacks from the Foolbox
art_tutorial	Run attacks from ART
attack_runner_tutorial	Organize a pentest.