Privacy preserving generative mechanism for industrial time-series data disclosure
Existing privacy-preserving techniques suffer from inherent drawbacks to retain characteristics of observed, and original industrial time series data for utility in the downstream tasks such as process modelling, control, optimization and etc. The embodiments herein provide a method and system for privacy preserving generative mechanism for data-disclosure of the industrial multivariate mixed-variable time series data. The system fuses an industrial time series data with a random gaussian noise to preserve the privacy of the industrial time series data and trades-off the privacy with the utility of synthetic-private data. Further, the system presents the privacy-preserving synthetic industrial data generative mechanism for data disclosure with minimal risk of AI technique and strong privacy guarantees. Embedding privacy by design into the generative mechanism approaches present an alternate paradigm of learning in contrast to the reduced-order modeling and numerical solutions of the industrial time-series data based on the principles in continuum mechanics for data disclosure with privacy.
1 . A processor-implemented method comprising steps of:
receiving, via an input/output interface, a multivariate mixed-variable time series data of a plurality of sensory observations, cluster-labels associated with the multivariate mixed-variable time series data, and a cluster-independent random noise, wherein multivariate mixed-variable time series data comprises continuous and discrete feature variables;
pre-processing, via one or more hardware processors, the received multivariate mixed-variable time series data, wherein the pre-processing comprising steps of:
normalizing the continuous feature variables by bounding heterogeneous measurements between a predefined range using a min-max scaling technique; and
transforming the discreate feature variables by representing as a sparse binary vector using a one-hot encoding technique,
training, via the one or more hardware processors, a plurality of neural networks of a privacy preserving generative adversarial network (ppGAN) in a first phase and a second phase using pre-processed multivariate mixed-variable time series data, wherein the plurality of neural networks including an embedding neural network, a recovery neural network, a generator neural network, a critic neural network and a discriminator neural network, wherein the discriminator neural network classifies the high-dimensional synthetic-private mixed feature embeddings as real or fake, wherein training the plurality of neural networks of the ppGAN in the first phase comprises:
training, via the one or more hardware processors, the embedding neural network using a predefined low-dimensional mixed feature training dataset to obtain a high-dimensional mixed feature embeddings, wherein the embedding neural network (E ppGAN ) comprises:
learning a high-dimensional representation,
H
train
n
,
1
:
T
n
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
by transforming the corresponding low-dimensional real sequences,
I
~
train
n
,
1
:
T
n
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
;
assisting in learning by incorporating semantics of the mixed-feature learning a high-dimensional representation,
H
train
n
,
1
:
T
n
;
and
E
ppGAN
:
I
train
n
,
1
:
T
n
∈
∏
t
∏
j
=
1
f
D
j
→
H
train
n
,
1
:
T
n
∈
∏
t
∏
j
=
1
f
H
j
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
wherein, D j , H j denotes a j-th feature variable space and a latent embedding vector space;
providing, via the one or more hardware processors, a test data to generate a synthetic private dataset for data disclosure using the trained plurality of neural networks of the ppGAN; and
estimating, via the one or more hardware processors, an identifiability of the multivariate mixed-variable time series data from the generated synthetic private dataset, wherein the estimation satisfies a predefined process-identifiability criteria.
2 . The processor-implemented method of claim 1 , wherein the first phase training of the plurality of neural networks of the ppGAN further comprising steps of:
training, via the one or more hardware processors, the recovery neural network using the obtained high-dimensional mixed feature embeddings to reconstruct the low-dimensional mixed feature dataset;
training, via the one or more hardware processors, the supervisor neural network using the obtained high-dimensional mixed-feature embeddings for a single step ahead predictions of the high-dimensional mixed-feature embeddings, wherein the supervisor neural network is utilized to model a temporal dynamics of the low-dimensional mixed feature training dataset, wherein trainable parameters of the embedding neural network and recovery neural network are updated by optimizing a cost function through the joint training of the networks in a supervised-learning approach by incorporating inductive bias of reconstruction of the original input by minimizing a supervised loss; and
training, via the one or more hardware processors, the critic neural network using the high-dimensional mixed feature embeddings to predict a target high-dimensional feature embedding, wherein the critic neural network is utilized to model the relationship between independent and dependent variables of the low-dimensional mixed feature training dataset, wherein the critic neural network comprises a neural-network function leveraged for predictive modeling using a predictive modeler that maps the independent mixed-feature variables to the target variable trained in the supervised learning approach and wherein the predictive modeler of the critic neural network is trained on a real data to map the independent feature variables to target variable by minimizing the prediction loss.
3 . The processor-implemented method of claim 1 , wherein a second phase training of the plurality of neural networks of a ppGAN comprising steps of:
transforming, via the one or more hardware processors, cluster-independent random noise using one or more cluster-labels associated with a predefined training dataset to obtain a cluster-dependent random noise;
performing, via the one or more hardware processors, a linear transformation on a concatenation of the low-dimensional mixed feature training dataset and the cluster-dependent random noise to obtain a synthetic-private noise;
training, via the one or more hardware processors, the generator neural network using the obtained synthetic-private noise to obtain a high-dimensional synthetic-private mixed feature embeddings;
training, via the one or more hardware processors, the critic neural network using the high-dimensional synthetic-private mixed feature embeddings to predict the synthetic-private target feature embedding;
training, via the one or more hardware processors, the discriminator neural network using the high-dimensional synthetic-private mixed feature embeddings to assign a label;
training, via the one or more hardware processors, the supervisory neural network using the high-dimensional synthetic-private mixed feature embeddings to generate a single-step ahead predictions of the high-dimensional synthetic-private mixed feature embeddings; and
training, via the one or more hardware processors, the recovery neural network using the single-step ahead high-dimensional synthetic-private mixed feature embeddings to obtain the low-dimensional synthetic-private mixed feature dataset.
4 . The processor-implemented method of claim 1 , wherein a low-dimensional mixed feature validation dataset is utilized for the hyper-parameter tuning of the ppGAN.
5 . A system comprising:
an input/output interface to a multivariate mixed-variable time series data of a plurality of sensory observations, cluster-labels associated with the multivariate mixed-variable time series data, and a cluster-independent random noise, wherein multivariate mixed-variable time series data comprises continuous and discrete feature variables;
a memory in communication with the one or more hardware processors, wherein the one or more hardware processors are configured to execute programmed instructions stored in the memory to:
pre-process the received multivariate mixed-variable time series data, wherein the pre-process includes normalizing the continuous feature variables by bounding heterogeneous measurements between a predefined range through a min-max scaling technique, and transforming discreate feature variables by representing as a sparse binary vector through a one-hot encoding technique;
train a plurality of neural networks of a privacy preserving generative adversarial network (ppGAN) in a first phase and a second phase using pre-processed multivariate mixed-variable time series data, wherein the plurality of neural networks including an embedding neural network, a recovery neural network, a generator neural network, a critic neural network, and a discriminator neural network, wherein the discriminator neural network classifies the high-dimensional synthetic-private mixed feature embeddings as real or fake, wherein training the plurality of neural networks of the ppGAN in the first phase comprises:
training, via the one or more hardware processors, the embedding neural network using a predefined low-dimensional mixed feature training dataset to obtain a high-dimensional mixed feature embeddings, wherein the embedding neural network (E ppGAN ) comprises:
learning a high-dimensional representation,
H
train
n
,
1
:
T
n
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
by transforming the corresponding low-dimensional real sequences,
I
~
train
n
,
1
:
T
n
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
;
assisting in learning by incorporating semantics of the mixed-feature variables in feature embeddings,
H
train
n
,
1
:
T
n
;
and
E
ppGAN
:
I
train
n
,
1
:
T
n
∈
∏
t
∏
j
=
1
f
D
j
→
H
train
n
,
1
:
T
n
∈
∏
t
∏
j
=
1
f
H
j
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
wherein, D j , H j denotes a j-th feature variable space and a latent embedding vector space;
provide a test data to generate a synthetic private dataset for data disclosure using the trained plurality of neural networks of the ppGAN; and
estimate an identifiability of the multivariate mixed-variable time series data from the generated synthetic private dataset, wherein the estimation satisfies a predefined process-identifiability criteria.
6 . The system of claim 5 , wherein a low-dimensional mixed feature validation dataset is utilized for the hyper-parameter tuning of the ppGAN.
7 . A non-transitory computer readable medium storing one or more instructions which when executed by one or more processors on a system, cause the one or more processors to perform method comprising steps of:
receiving, via an input/output interface, a multivariate mixed-variable time series data of a plurality of sensory observations, cluster-labels associated with the multivariate mixed-variable time series data, and a cluster-independent random noise, wherein multivariate mixed-variable time series data comprises continuous and discrete feature variables;
pre-processing, via one or more hardware processors, the received multivariate mixed-variable time series data, wherein the pre-processing comprising steps of:
normalizing the continuous feature variables by bounding heterogeneous measurements between a predefined range through a min-max scaling technique; and
transforming the discreate feature variables by representing as a sparse binary vector using a one-hot encoding technique,
training, via the one or more hardware processors, a plurality of neural networks of a privacy preserving generative adversarial network (ppGAN) in a first phase and a second phase using pre-processed multivariate mixed-variable time series data, wherein the plurality of neural networks including an embedding neural network, a recovery neural network, a generator neural network, a critic neural network, and a discriminator neural network, wherein the discriminator neural network classifies the high-dimensional synthetic-private mixed feature embeddings as real or fake, wherein training the plurality of neural networks of the ppGAN in the first phase comprises:
training, via the one or more hardware processors, the embedding neural network using a predefined low-dimensional mixed feature training dataset to obtain a high-dimensional mixed feature embeddings, wherein the embedding neural network (E ppGAN ) comprises:
learning a high-dimensional representation,
H
train
n
,
1
:
T
n
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
by transforming the corresponding low-dimensional real sequences,
I
~
train
n
,
1
:
T
n
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
;
{(1, 2, . . . , N)};
assisting in learning by incorporating semantics of the mixed-feature variables in feature embeddings,
H
train
n
,
1
:
T
n
;
and
E
ppGAN
:
I
train
n
,
1
:
T
n
∈
∏
t
∏
j
=
1
f
D
j
→
H
train
n
,
1
:
T
n
∈
∏
t
∏
j
=
1
f
H
j
,
∀
n
∈
{
(
1
,
2
,
…
,
N
)
}
wherein, D j , H j denotes a j-th feature variable space and a latent embedding vector space;
providing, via the one or more hardware processors, a test data to generate a synthetic private dataset for data disclosure using the trained plurality of neural networks of the ppGAN; and
estimating, via the one or more hardware processors, an identifiability of the multivariate mixed-variable time series data from the generated synthetic private dataset, wherein the estimation satisfies a predefined process-identifiability criteria.