anomed-deanonymizer 0.0.7

A library aiding to create deanonymizers (attacks on privacy preserving machine learning models) for the AnoMed competition platform.

Deanonymizer

A library aiding to create attacks against anonymizers (privacy preserving machine learning models) for the AnoMed competition platform. Currently, only membership inference attacks are supported.

Usage Example

The following example will create a Falcon-based web app that encapsulates a deanonymizer, targeting the example anonymizer defined in the anomed-anonyimzer README.md (which is a privacy preserving classifier for the famous iris dataset classification problem). The encapsulated deanonymizer is a membership inference black box attack, implemented using the Adversarial Robustness Toolbox (ART library).

The web app offers these routes (some may have query parameters not mentioned here):

[GET] / (This displays an "alive message".)
[POST] /fit (This invokes fitting the Gaussian naive based classifier; the web app will pull the training data from training_data_url.)
[POST] /evaluate (This invokes an intermediate, or final evaluation of the classifier.)

import anomed_deanonymizer
import numpy as np
from art.attacks.inference.membership_inference import MembershipInferenceBlackBox


def validate_input_array(feature_array: np.ndarray) -> None:
    if feature_array.shape[1] != 4 or len(feature_array.shape) != 2:
        raise ValueError("Feature array needs to have shape (n_samples, 4).")
    if feature_array.dtype != np.float_:
        raise ValueError("Feature array must be an array of floats.")


attack_target = anomed_deanonymizer.WebClassifier(
    url="http://example.com/predict", input_shape=(4,), nb_classes=3
)
example_attack_art = MembershipInferenceBlackBox(estimator=attack_target)  # type: ignore
example_attack = anomed_deanonymizer.ARTWrapper(
    art_mia=example_attack_art, input_validator=validate_input_array
)
application = anomed_deanonymizer.supervised_learning_MIA_server_factory(
    anonymizer_identifier="example_anonymizer",
    deanonymizer_identifier="example_deanonymizer",
    deanonymizer_obj=example_attack,
    model_filepath="deanonymizer.pkl",
    default_batch_size=64,
    member_url="http://example.com/members",
    nonmember_url="http://example.com/non-members",
    evaluation_data_url="http://example.com/attack-success-evaluation",
    model_loader=anomed_deanonymizer.unpickle_deanonymizer,
    utility_evaluation_url="http://example.com/utility",
)