Method and system for identifying mislabeled data samples using adversarial attacks

1 . A processor implemented method of identifying mislabeled data samples, comprising:

collecting training data, via one or more hardware processors, wherein the training data comprises the mislabeled data samples and correctly labelled data samples, wherein the labeling is done by a human or by a system, wherein the data samples are real-time data from data sources, wherein the data sources comprises industrial plants, and/or artificially generated using soft sensors or laboratory experiments, wherein the mislabeled sample comprises a) a data sample that is tagged with a wrong/incorrect label, or b) a distorted/confusing data sample having similarity with multiple classes;

training a data-driven model using the training data, via the one or more hardware processors;

computing, using trained data-driven model, logit or probability scores corresponding to a plurality of data samples in the training data, via the one or more hardware processors;

performing an adversarial attack on each of the plurality of data samples in the training data, via the one or more hardware processors, wherein the adversarial attack is performed by considering a data label predicted by the data-driven model as a true label of each data sample, wherein the adversarial attack is performed to obtain wrong result from a machine learning model or the data-driven model, by designing an input in a way using varying values of adversarial perturbations, wherein performing the adversarial attack, comprising:

identifying all data samples from among the plurality of data samples, that have been misguided by the adversarial attack;

identifying for each of the data samples that have been misguided by the adversarial attack, least value of adversarial perturbation from among a plurality of values of adversarial perturbations used for performing the adversarial attack on each of the plurality of data samples, in terms of a predefined metric representing strength of the adversarial attack, wherein the predefined metric is computed using an original data sample and corresponding adversarial data sample, for each of the data samples identified as misguided by the adversarial attack; and

sorting the data samples that have been misguided by the adversarial attack, in ascending order of values of the predefined metric; and

identifying candidate mislabeled data samples from among the plurality of data samples, via the one or more hardware processors, wherein data samples for which the value of predefined metric is below a predefined threshold are identified as the candidate mislabeled data samples, thereby improving accuracy of identification of the mislabel data samples.

2 . The method as claimed in claim 1 , wherein the predefined metric is dependent on data type and is at least one of Dynamic Time Warping Distance for time series data, perceptual loss for image and video data, and Euclidean distance for tabular data.

3 . A system of identifying mislabeled data samples, comprising:

one or more hardware processors;

a communication interface; and

a memory storing a plurality of instructions, wherein the plurality of instructions when executed, cause the one or more hardware processors to:

collect training data, wherein the training data comprises the mislabeled data samples and correctly labelled data samples, wherein the labeling is done by a human or by a system, wherein the data samples are real-time data from data sources, wherein the data sources comprises industrial plants, and/or artificially generated using soft sensors or laboratory experiments, wherein the mislabeled sample comprises a) a data sample that is tagged with a wrong/incorrect label, or b) a distorted/confusing data sample having similarity with multiple classes;

train a data-driven model using the training data;

compute, using trained data-driven model, logit or probability scores corresponding to a plurality of data samples in the training data;

perform adversarial attack on each of the plurality of data samples in the training data, wherein the adversarial attack is performed by considering a data label predicted by the data-driven model as a true label of each data sample, wherein the adversarial attack is performed to obtain wrong result from a machine learning model or the data-driven model, by designing an input in a way using varying values of adversarial perturbations, wherein performing the adversarial attack, by:

identifying all data samples from among the plurality of data samples, that have been misguided by the adversarial attack;

identifying for each of the data samples that have been misguided by the adversarial attack, least value of adversarial perturbation from among a plurality of values of adversarial perturbations used for performing the adversarial attack on each of the plurality of data samples, in terms of a predefined metric representing strength of the adversarial attack; and

sorting the data samples that have been misguided by the adversarial attack, in ascending order of values of the predefined metric; and

identify candidate mislabeled data samples from among the plurality of data samples, wherein data samples for which the value of predefined metric is below a predefined threshold are identified as the candidate mislabeled data samples, thereby improving accuracy of identification of the mislabel data samples.

4 . The system as claimed in claim 3 , wherein the predefined metric is dependent on data type and is at least one of Dynamic Time Warping Distance for time series data, perceptual loss for image and video data, and Euclidean distance for tabular data.

5 . One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:

collecting training data, wherein the training data comprises the mislabeled data samples and correctly labelled data samples, wherein the labeling is done by a human or by a system, wherein the data samples are real-time data from data sources, wherein the data sources comprises industrial plants, and/or artificially generated using soft sensors or laboratory experiments, wherein the mislabeled sample comprises a) a data sample that is tagged with a wrong/incorrect label, or b) a distorted/confusing data sample having similarity with multiple classes;

training a data-driven model using the training data;

computing, using trained data-driven model, logit or probability scores corresponding to a plurality of data samples in the training data;

performing an adversarial attack on each of the plurality of data samples in the training data, wherein the adversarial attack is performed by considering a data label predicted by the data-driven model as a true label of each data sample, wherein the adversarial attack is performed to obtain wrong result from a machine learning model or the data-driven model, by designing an input in a way using varying values of adversarial perturbations, wherein performing the adversarial attack comprising:

identifying all data samples from among the plurality of data samples, that have been misguided by the adversarial attack;

identifying for each of the data samples that have been misguided by the adversarial attack, least value of adversarial perturbation from among a plurality of values of adversarial perturbations used for performing the adversarial attack on each of the plurality of data samples, in terms of a predefined metric representing strength of the adversarial attack; and

sorting the data samples that have been misguided by the adversarial attack, in ascending order of values of the predefined metric; and

identifying candidate mislabeled data samples from among the plurality of data samples, wherein data samples for which the value of predefined metric is below a predefined threshold are identified as the candidate mislabeled data samples, thereby improving accuracy of identification of the mislabel data samples.

6 . The one or more non-transitory machine-readable information storage mediums of claim 5 , wherein the predefined metric is dependent on data type and is at least one of Dynamic Time Warping Distance for time series data, perceptual loss for image and video data, and Euclidean distance for tabular data.