mindarmour.privacy.evaluation

This module provides some methods to evaluate the risk of privacy leakage of given model.

class mindarmour.privacy.evaluation.ImageInversionAttack(network, input_shape, input_bound, loss_weights=(1, 0.2, 5))[source]

An attack method used to reconstruct images by inverting their deep representations.

References: Aravindh Mahendran, Andrea Vedaldi. Understanding Deep Image Representations by Inverting Them. 2014.

Parameters

  • network (Cell) – The network used to infer images’ deep representations.

  • input_shape (tuple) – Data shape of single network input, which should be in accordance with the given network. The format of shape should be (channel, image_width, image_height).

  • input_bound (Union[tuple, list]) – The pixel range of original images, which should be like [minimum_pixel, maximum_pixel] or (minimum_pixel, maximum_pixel).

  • loss_weights (Union[list, tuple]) – Weights of three sub-loss in InversionLoss, which can be adjusted to obtain better results. Default: (1, 0.2, 5).

Raises

  • TypeError – If the type of network is not Cell.

  • ValueError – If any value of input_shape is not positive int.

  • ValueError – If any value of loss_weights is not positive value.

Examples

>>> import mindspore.ops.operations as P
>>> from mindspore.nn import Cell
>>> from mindarmour.privacy.evaluation.inversion_attack import ImageInversionAttack
>>> class Net(Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self._softmax = P.Softmax()
...         self._reduce = P.ReduceSum()
...         self._squeeze = P.Squeeze(1)
...     def construct(self, inputs):
...         out = self._softmax(inputs)
...         out = self._reduce(out, 2)
...         return self._squeeze(out)
>>> net = Net()
>>> original_images = np.random.random((2,1,10,10)).astype(np.float32)
>>> target_features =  np.random.random((2,10)).astype(np.float32)
>>> inversion_attack = ImageInversionAttack(net,
...                                         input_shape=(1, 10, 10),
...                                         input_bound=(0, 1),
...                                         loss_weights=[1, 0.2, 5])
>>> inversion_images = inversion_attack.generate(target_features, iters=10)
>>> evaluate_result = inversion_attack.evaluate(original_images, inversion_images)
evaluate(original_images, inversion_images, labels=None, new_network=None)[source]

Evaluate the quality of inverted images by three index: the average L2 distance and SSIM value between original images and inversion images, and the average of inverted images’ confidence on true labels of inverted inferred by a new trained network.

Parameters

  • original_images (numpy.ndarray) – Original images, whose shape should be (img_num, channels, img_width, img_height).

  • inversion_images (numpy.ndarray) – Inversion images, whose shape should be (img_num, channels, img_width, img_height).

  • labels (numpy.ndarray) – Ground truth labels of original images. Default: None.

  • new_network (Cell) – A network whose structure contains all parts of self._network, but loaded with different checkpoint file. Default: None.

Returns

  • float, l2 distance.

  • float, average ssim value.

  • Union[float, None], average confidence. It would be None if labels or new_network is None.

generate(target_features, iters=100)[source]

Reconstruct images based on target_features.

Parameters

  • target_features (numpy.ndarray) – Deep representations of original images. The first dimension of target_features should be img_num. It should be noted that the shape of target_features should be (1, dim2, dim3, …) if img_num equals 1.

  • iters (int) – iteration times of inversion attack, which should be positive integers. Default: 100.

Returns

numpy.ndarray, reconstructed images, which are expected to be similar to original images.

Raises

  • TypeError – If the type of target_features is not numpy.ndarray.

  • ValueError – If any value of iters is not positive int.Z

class mindarmour.privacy.evaluation.MembershipInference(model, n_jobs=- 1)[source]

Proposed by Shokri, Stronati, Song and Shmatikov, membership inference is a grey-box attack for inferring user’s privacy data. It requires loss or logits results of the training samples. (Privacy refers to some sensitive attributes of a single user).

For details, please refer to the Tutorial.

References: Reza Shokri, Marco Stronati, Congzheng Song, Vitaly Shmatikov. Membership Inference Attacks against Machine Learning Models. 2017.

Parameters

  • model (Model) – Target model.

  • n_jobs (int) – Number of jobs run in parallel. -1 means using all processors, otherwise the value of n_jobs must be a positive integer.

Raises

  • TypeError – If type of model is not mindspore.train.Model.

  • TypeError – If type of n_jobs is not int.

  • ValueError – The value of n_jobs is neither -1 nor a positive integer.

Examples

>>> import mindspore.ops.operations as P
>>> from mindspore.nn import Cell
>>> from mindspore import Model
>>> from mindarmour.privacy.evaluation import MembershipInference
>>> def dataset_generator():
...     batch_size = 16
...     batches = 1
...     data =  np.random.randn(batches * batch_size,1,10).astype(np.float32)
...     label =  np.random.randint(0,10, batches * batch_size).astype(np.int32)
...     for i in range(batches):
...         yield data[i*batch_size:(i+1)*batch_size], label[i*batch_size:(i+1)*batch_size]
>>> class Net(Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self._softmax = P.Softmax()
...         self._Dense = nn.Dense(10,10)
...         self._squeeze = P.Squeeze(1)
...     def construct(self, inputs):
...         out = self._softmax(inputs)
...         out = self._Dense(out)
...         return self._squeeze(out)
>>> net = Net()
>>> loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True)
>>> opt = nn.Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> model = Model(network=net, loss_fn=loss, optimizer=opt)
>>> inference_model = MembershipInference(model, 2)
>>> config = [{
...     "method": "KNN",
...     "params": {"n_neighbors": [3, 5, 7],}
...     }]
>>> ds_train = ds.GeneratorDataset(dataset_generator, ["image", "label"])
>>> ds_test = ds.GeneratorDataset(dataset_generator, ["image", "label"])
>>> inference_model.train(ds_train, ds_test, config)
>>> metrics = ["precision", "accuracy", "recall"]
>>> eval_train = ds.GeneratorDataset(dataset_generator, ["image", "label"])
>>> eval_test = ds.GeneratorDataset(dataset_generator, ["image", "label"])
>>> result = inference_model.eval(eval_train. eval_test, metrics)
>>> print(result)
eval(dataset_train, dataset_test, metrics)[source]

Evaluate the different privacy of the target model. Evaluation indicators shall be specified by metrics.

Parameters

  • dataset_train (mindspore.dataset) – The training dataset for the target model.

  • dataset_test (mindspore.dataset) – The test dataset for the target model.

  • metrics (Union[list, tuple]) – Evaluation indicators. The value of metrics must be in [“precision”, “accuracy”, “recall”]. Default: [“precision”].

Returns

list, each element contains an evaluation indicator for the attack model.

train(dataset_train, dataset_test, attack_config)[source]

Depending on the configuration, use the input dataset to train the attack model. Save the attack model to self._attack_list.

Parameters

  • dataset_train (mindspore.dataset) – The training dataset for the target model.

  • dataset_test (mindspore.dataset) – The test set for the target model.

  • attack_config (Union[list, tuple]) – Parameter setting for the attack model. The format is [{“method”: “knn”, “params”: {“n_neighbors”: [3, 5, 7]}}, {“method”: “lr”, “params”: {“C”: np.logspace(-4, 2, 10)}}]. The support methods are knn, lr, mlp and rf, and the params of each method must within the range of changeable parameters. Tips of params implement can be found below: KNN, LR, RF, MLP.

Raises

  • KeyError – If any config in attack_config doesn’t have keys {“method”, “params”}.

  • NameError – If the method(case insensitive) in attack_config is not in [“lr”, “knn”, “rf”, “mlp”].