Heterogeneous Neural Networks

Neural networks are probably the most popular machine learning algorithms in recent years. FATE provides a federated Heterogeneous neural network implementation.

This federated heterogeneous neural network framework allows multiple parties to jointly conduct a learning process with partially overlapping user samples but different feature sets, which corresponds to a vertically partitioned virtual data set. An advantage of Hetero NN is that it provides the same level of accuracy as the non privacy-preserving approach while at the same time, reveal no information of each private data provider.

Basic FrameWork

The following figure shows the proposed Federated Heterogeneous Neural Network framework.

Party B: We define the party B as the data provider who holds both a data matrix and the class label. Since the class label information is indispensable for supervised learning, there must be an party with access to the label y. The party B naturally takes the responsibility as a dominating server in federated learning.

Party A: We define the data provider which has only a data matrix as party A. Party A plays the role of clients in the federated learning setting.

The data samples are aligned under an encryption scheme. By using the privacy-preserving protocol for inter-database intersections, the parties can find their common users or data samples without compromising the non-overlapping parts of the data sets.

Party B and party A each have their own bottom neural network model, which may be different. The parties jointly build the interactive layer, which is a fully connected layer. This layer's input is the concatenation of the two parties' bottom model output. In addition, only party B owns the model of interactive layer. Lastly, party B builds the top neural network model and feeds the output of interactive layer to it.

Forward Propagation of Federated Heterogeneous Neural Network

Forward Propagation Process consists of three parts.

• Part Ⅰ
  Forward Propagation of Bottom Model.

1. Party A feeds its input features X to its bottom model and gets the forward output of bottom model alpha_A
2. Party B feeds its input features X to its bottom model and gets the forward output of bottom model alpha_B if active party has input features.

• Part ⅠⅠ
  Forward Propagation of Interactive Layer.

1. Party A uses additive homomorphic encryption to encrypt alpha_A(mark as $$alpha \_A$$ ), and sends the encrypted result to party B.
2. Party B receives the $$alpha\_A$$ , multiplies it by interactive layer's party A model weight W_A, get $$z\_A$$ . Party B also multiplies its interactive layer's weight W_B by its own bottom output, getting z_B. Party B generates noise epsilon_B, adds it to $$z\_A$$ and sends addition result to party A.
3. Party A calculates the product of accumulate noise epsilon_acc and bottom input alpha_A (epsilon_acc * alpha_A). Decrypting the received result $$z\_A + epsilon\_B$$ , Party A adds the product to it and sends result to Active party.
4. Party B subtracts the party A's sending value by epsilon_B( get z_A + epsilon_acc * alpha_A), and feeds z = z_A + epsilon_acc * alpha_A + z_B(if exists) to activation function.

• Part ⅠⅠⅠ
  Forward Propagation of Top Model.

1. Party B takes the output of activation function's output of interactive layer g(z) and runs the forward process of top model. The following figure shows the forward propagation of Federated Heterogeneous Neural Network framework.

Backward Propagation of Federated Heterogeneous Neural Network

Backward Propagation Process also consists of three parts.

• Part I
  Backward Propagation of Top Model.

1. Party B calculates the error delta of interactive layer output, then updates top model.

• Part II
  Backward Propagation of Interactive layer.

1. Party B calculates the error delta_act of activation function's output by delta.
2. Party B propagates delta_bottomB = delta_act * W_B to bottom model, then updates W_B(W_B -= eta * delta_act * alpha_B).
3. Party B generates noise epsilon_B, calculates $$delta\_act \* (alpha\_A + epsilon\_B$$ and sends it to party A.
4. Party A encrypts epsilon_acc, sends $$epsilon\_acc$$ to party B. Then party B decrypts the received value. Party A generates noise epsilon_A, adds epsilon_A / eta to decrypted result(delta_act * alpha_A + epsilon_B + epsilon_A / eta) and add epsilon_A to accumulate noise epsilon_acc(epsilon_acc += epsilon_A). Party A sends the addition result to party B. (delta_act * W_A + epsilon_B + epsilon_A / eta)
5. Party B receives $$epsilon\_acc$$ and delta_act * alpha_A + epsilon_B + epsilon_A / eta. Firstly it sends party A's bottom model output' error $$delta\_act \* W\_A + acc$$ to party A. Secondly updates W_A -= eta * (delta_act * W_A + epsilon_B + epsilon_A / eta - epsilon_B) = eta * delta_act * W_A - epsilon_B = W_TRUE - epsilon_acc. Where W_TRUE represents the actually weights.
6. Party A decrypts $$delta\_act \* (W\_A + acc)$$ and passes delta_act * (W_A + acc) to its bottom model.

• Part III
  Backward Propagation of Bottom Model.

1. Party B and party A updates their bottom model separately. The following figure shows the backward propagation of Federated Heterogeneous Neural Network framework.

Param

hetero_nn_param

Classes

DatasetParam(dataset_name=None, **kwargs)

Bases: BaseParam

Source code in python/federatedml/param/hetero_nn_param.py

def __init__(self, dataset_name=None, **kwargs):
    super(DatasetParam, self).__init__()
    self.dataset_name = dataset_name
    self.param = kwargs

Attributes

dataset_name = dataset_name instance-attribute

param = kwargs instance-attribute

Functions

check()

Source code in python/federatedml/param/hetero_nn_param.py

def check(self):
    if self.dataset_name is not None:
        self.check_string(self.dataset_name, 'dataset_name')

to_dict()

Source code in python/federatedml/param/hetero_nn_param.py

def to_dict(self):
    ret = {'dataset_name': self.dataset_name, 'param': self.param}
    return ret

SelectorParam(method=None, beta=1, selective_size=consts.SELECTIVE_SIZE, min_prob=0, random_state=None)

Bases: object

Parameters:

Name	Type	Description	Default
method		back propagation select method, accept "relative" only, default: None	None
selective_size		deque size to use, store the most recent selective_size historical loss, default: 1024	consts.SELECTIVE_SIZE
beta		sample whose selective probability >= power(np.random, beta) will be selected	1
min_prob		selective probability is max(min_prob, rank_rate)	0

Source code in python/federatedml/param/hetero_nn_param.py

def __init__(self, method=None, beta=1, selective_size=consts.SELECTIVE_SIZE, min_prob=0, random_state=None):
    self.method = method
    self.selective_size = selective_size
    self.beta = beta
    self.min_prob = min_prob
    self.random_state = random_state

Attributes

method = method instance-attribute

selective_size = selective_size instance-attribute

beta = beta instance-attribute

min_prob = min_prob instance-attribute

random_state = random_state instance-attribute

Functions

check()

Source code in python/federatedml/param/hetero_nn_param.py

def check(self):
    if self.method is not None and self.method not in ["relative"]:
        raise ValueError('selective method should be None be "relative"')

    if not isinstance(self.selective_size, int) or self.selective_size <= 0:
        raise ValueError("selective size should be a positive integer")

    if not isinstance(self.beta, int):
        raise ValueError("beta should be integer")

    if not isinstance(self.min_prob, (float, int)):
        raise ValueError("min_prob should be numeric")

CoAEConfuserParam(enable=False, epoch=50, lr=0.001, lambda1=1.0, lambda2=2.0, verbose=False)

Bases: BaseParam

A label protect mechanism proposed in paper: "Batch Label Inference and Replacement Attacks in Black-Boxed Vertical Federated Learning" paper link: https://arxiv.org/abs/2112.05409 Convert true labels to fake soft labels by using an auto-encoder.

Args: enable: boolean run CoAE or not epoch: None or int auto-encoder training epochs lr: float auto-encoder learning rate lambda1: float parameter to control the difference between true labels and fake soft labels. Larger the parameter, autoencoder will give more attention to making true labels and fake soft label different. lambda2: float parameter to control entropy loss, see original paper for details verbose: boolean print loss log while training auto encoder

Source code in python/federatedml/param/hetero_nn_param.py

def __init__(self, enable=False, epoch=50, lr=0.001, lambda1=1.0, lambda2=2.0, verbose=False):
    super(CoAEConfuserParam, self).__init__()
    self.enable = enable
    self.epoch = epoch
    self.lr = lr
    self.lambda1 = lambda1
    self.lambda2 = lambda2
    self.verbose = verbose

Attributes

enable = enable instance-attribute

epoch = epoch instance-attribute

lr = lr instance-attribute

lambda1 = lambda1 instance-attribute

lambda2 = lambda2 instance-attribute

verbose = verbose instance-attribute

Functions

check()

Source code in python/federatedml/param/hetero_nn_param.py

def check(self):

    self.check_boolean(self.enable, 'enable')

    if not isinstance(self.epoch, int) or self.epoch <= 0:
        raise ValueError("epoch should be a positive integer")

    if not isinstance(self.lr, float):
        raise ValueError('lr should be a float number')

    if not isinstance(self.lambda1, float):
        raise ValueError('lambda1 should be a float number')

    if not isinstance(self.lambda2, float):
        raise ValueError('lambda2 should be a float number')

    self.check_boolean(self.verbose, 'verbose')

HeteroNNParam(task_type='classification', bottom_nn_define=None, top_nn_define=None, interactive_layer_define=None, interactive_layer_lr=0.9, config_type='pytorch', optimizer='SGD', loss=None, epochs=100, batch_size=-1, early_stop='diff', tol=1e-05, seed=100, encrypt_param=EncryptParam(), encrypted_mode_calculator_param=EncryptedModeCalculatorParam(), predict_param=PredictParam(), cv_param=CrossValidationParam(), validation_freqs=None, early_stopping_rounds=None, metrics=None, use_first_metric_only=True, selector_param=SelectorParam(), floating_point_precision=23, callback_param=CallbackParam(), coae_param=CoAEConfuserParam(), dataset=DatasetParam())

Bases: BaseParam

Parameters used for Hetero Neural Network.

Parameters:

Name	Type	Description	Default
task_type			'classification'
bottom_nn_define			None
interactive_layer_define			None
interactive_layer_lr			0.9
top_nn_define			None
optimizer		a string, one of "Adadelta", "Adagrad", "Adam", "Adamax", "Nadam", "RMSprop", "SGD" a dict, with a required key-value pair keyed by "optimizer", with optional key-value pairs such as learning rate. defaults to "SGD".	'SGD'
loss			None
epochs			100
batch_size	int, batch size when updating model.	-1 means use all data in a batch. i.e. Not to use mini-batch strategy. defaults to -1.	-1
early_stop	str, accept 'diff' only in this version, default	Method used to judge converge or not. a) diff： Use difference of loss between two iterations to judge whether converge.	'diff'
floating_point_precision		e.g.: convert an x to round(x * 2**floating_point_precision) during Paillier operation, divide the result by 2**floating_point_precision in the end.	23
callback_param			CallbackParam()

Source code in python/federatedml/param/hetero_nn_param.py

def __init__(self,
             task_type='classification',
             bottom_nn_define=None,
             top_nn_define=None,
             interactive_layer_define=None,
             interactive_layer_lr=0.9,
             config_type='pytorch',
             optimizer='SGD',
             loss=None,
             epochs=100,
             batch_size=-1,
             early_stop="diff",
             tol=1e-5,
             seed=100,
             encrypt_param=EncryptParam(),
             encrypted_mode_calculator_param=EncryptedModeCalculatorParam(),
             predict_param=PredictParam(),
             cv_param=CrossValidationParam(),
             validation_freqs=None,
             early_stopping_rounds=None,
             metrics=None,
             use_first_metric_only=True,
             selector_param=SelectorParam(),
             floating_point_precision=23,
             callback_param=CallbackParam(),
             coae_param=CoAEConfuserParam(),
             dataset=DatasetParam()
             ):

    super(HeteroNNParam, self).__init__()

    self.task_type = task_type
    self.bottom_nn_define = bottom_nn_define
    self.interactive_layer_define = interactive_layer_define
    self.interactive_layer_lr = interactive_layer_lr
    self.top_nn_define = top_nn_define
    self.batch_size = batch_size
    self.epochs = epochs
    self.early_stop = early_stop
    self.tol = tol
    self.optimizer = optimizer
    self.loss = loss
    self.validation_freqs = validation_freqs
    self.early_stopping_rounds = early_stopping_rounds
    self.metrics = metrics or []
    self.use_first_metric_only = use_first_metric_only
    self.encrypt_param = copy.deepcopy(encrypt_param)
    self.encrypted_model_calculator_param = encrypted_mode_calculator_param
    self.predict_param = copy.deepcopy(predict_param)
    self.cv_param = copy.deepcopy(cv_param)
    self.selector_param = selector_param
    self.floating_point_precision = floating_point_precision
    self.callback_param = copy.deepcopy(callback_param)
    self.coae_param = coae_param
    self.dataset = dataset
    self.seed = seed
    self.config_type = 'pytorch'  # pytorch only

Attributes

task_type = task_type instance-attribute

bottom_nn_define = bottom_nn_define instance-attribute

interactive_layer_define = interactive_layer_define instance-attribute

interactive_layer_lr = interactive_layer_lr instance-attribute

top_nn_define = top_nn_define instance-attribute

batch_size = batch_size instance-attribute

epochs = epochs instance-attribute

early_stop = early_stop instance-attribute

tol = tol instance-attribute

optimizer = optimizer instance-attribute

loss = loss instance-attribute

validation_freqs = validation_freqs instance-attribute

early_stopping_rounds = early_stopping_rounds instance-attribute

metrics = metrics or [] instance-attribute

use_first_metric_only = use_first_metric_only instance-attribute

encrypt_param = copy.deepcopy(encrypt_param) instance-attribute

encrypted_model_calculator_param = encrypted_mode_calculator_param instance-attribute

predict_param = copy.deepcopy(predict_param) instance-attribute

cv_param = copy.deepcopy(cv_param) instance-attribute

selector_param = selector_param instance-attribute

floating_point_precision = floating_point_precision instance-attribute

callback_param = copy.deepcopy(callback_param) instance-attribute

coae_param = coae_param instance-attribute

dataset = dataset instance-attribute

seed = seed instance-attribute

config_type = 'pytorch' instance-attribute

Functions

check()

Source code in python/federatedml/param/hetero_nn_param.py

def check(self):

    assert isinstance(self.dataset, DatasetParam), 'dataset must be a DatasetParam()'

    self.dataset.check()

    self.check_positive_integer(self.seed, 'seed')

    if self.task_type not in ["classification", "regression"]:
        raise ValueError("config_type should be classification or regression")

    if not isinstance(self.tol, (int, float)):
        raise ValueError("tol should be numeric")

    if not isinstance(self.epochs, int) or self.epochs <= 0:
        raise ValueError("epochs should be a positive integer")

    if self.bottom_nn_define and not isinstance(self.bottom_nn_define, dict):
        raise ValueError("bottom_nn_define should be a dict defining the structure of neural network")

    if self.top_nn_define and not isinstance(self.top_nn_define, dict):
        raise ValueError("top_nn_define should be a dict defining the structure of neural network")

    if self.interactive_layer_define is not None and not isinstance(self.interactive_layer_define, dict):
        raise ValueError(
            "the interactive_layer_define should be a dict defining the structure of interactive layer")

    if self.batch_size != -1:
        if not isinstance(self.batch_size, int) \
                or self.batch_size < consts.MIN_BATCH_SIZE:
            raise ValueError(
                " {} not supported, should be larger than 10 or -1 represent for all data".format(self.batch_size))

    if self.early_stop != "diff":
        raise ValueError("early stop should be diff in this version")

    if self.metrics is not None and not isinstance(self.metrics, list):
        raise ValueError("metrics should be a list")

    if self.floating_point_precision is not None and \
            (not isinstance(self.floating_point_precision, int) or
             self.floating_point_precision < 0 or self.floating_point_precision > 63):
        raise ValueError("floating point precision should be null or a integer between 0 and 63")

    self.encrypt_param.check()
    self.encrypted_model_calculator_param.check()
    self.predict_param.check()
    self.selector_param.check()
    self.coae_param.check()

    descr = "hetero nn param's "

    for p in ["early_stopping_rounds", "validation_freqs",
              "use_first_metric_only"]:
        if self._deprecated_params_set.get(p):
            if "callback_param" in self.get_user_feeded():
                raise ValueError(f"{p} and callback param should not be set simultaneously，"
                                 f"{self._deprecated_params_set}, {self.get_user_feeded()}")
            else:
                self.callback_param.callbacks = ["PerformanceEvaluate"]
            break

    if self._warn_to_deprecate_param("validation_freqs", descr, "callback_param's 'validation_freqs'"):
        self.callback_param.validation_freqs = self.validation_freqs

    if self._warn_to_deprecate_param("early_stopping_rounds", descr, "callback_param's 'early_stopping_rounds'"):
        self.callback_param.early_stopping_rounds = self.early_stopping_rounds

    if self._warn_to_deprecate_param("metrics", descr, "callback_param's 'metrics'"):
        if self.metrics:
            self.callback_param.metrics = self.metrics

    if self._warn_to_deprecate_param("use_first_metric_only", descr, "callback_param's 'use_first_metric_only'"):
        self.callback_param.use_first_metric_only = self.use_first_metric_only

Functions

Other features

• Allow party B's training without features.
• Support evaluate training and validate data during training process
• Support use early stopping strategy since FATE-v1.4.0
• Support selective backpropagation since FATE-v1.6.0
• Support low floating-point optimization since FATE-v1.6.0
• Support drop out strategy of interactive layer since FATE-v1.6.0

[1] Zhang Q, Wang C, Wu H, et al. GELU-Net: A Globally Encrypted, Locally Unencrypted Deep Neural Network for Privacy-Preserved Learning $$C$$ //IJCAI. 2018: 3933-3939.

[2] Zhang Y, Zhu H. Additively Homomorphical Encryption based Deep Neural Network for Asymmetrically Collaborative Machine Learning $$J$$ . arXiv preprint arXiv:2007.06849, 2020.

---

Last update: 2021-11-15