Model shift prevention through machine learning

1 . A method comprising:

receiving, by a computer, a set of input data associated with an entity having a credential allowing the entity to communicate with the computer;

applying, by the computer as an input, the set of input data to a trained current machine learning model;

in response to the input of the set of input data, producing, by the trained current machine learning model, a first set of classification data;

updating, by the computer, the trained current machine learning model using the set of input data and the first set of classification data to generate an updated current machine learning model;

determining, by the computer, first metadata corresponding to the trained current machine learning model before the updating, the first metadata corresponding to a metric or a characteristic of the trained current machine learning model before the updating;

determining, by the computer, second metadata corresponding to the updated current machine learning model, the second metadata corresponding to a metric or a characteristic of the updated current machine learning model;

analyzing, by the computer, the first metadata and the second metadata to determine a deviation metric;

determining, by the computer, that the deviation metric is greater than a deviation threshold;

in response to the deviation metric being greater than the deviation threshold:

applying, by the computer, the set of input data to a plurality of previously generated machine learning models, respectively, wherein the plurality of previously generated machine learning models do not include the updated current machine learning model;

producing, by the plurality of previously generated machine learning models, a plurality of sets of classification data in response to respectively applying the set of input data to the plurality of previously generated machine learning models;

determining, by the computer, whether the entity associated with the set of input data is a malicious entity attempting to induce a malicious shift of the trained current machine learning model toward incorrect classifications, the determining comprising:

combining the plurality of sets of classification data to produce a combined set of classification data, wherein both the combined set of classification data and the first classification set of data are produced for a same set of input data;

analyzing the combined set of classification data and the first set of classification data to determine an error metric, wherein the error metric indicates a difference in (1) one or more classifications included in the first set of classification data produced by the updated current machine learning model and (2) classifications included in the combined set of classification data from the plurality of previously generated machine learning models, wherein the error metric is a measure of a classification divergence between the combined set of classification data and the first classification set of data;

determining that the error metric is greater than an error threshold; and

in response to the error metric being greater than the error threshold, determining that the entity associated with the set of input data is the malicious entity that is inducing the malicious shift of the trained current machine learning model toward the incorrect classifications; and

responsive to the determining that the entity is the malicious entity:

retraining, by the computer, the updated current machine learning model by providing, as an input, the set of input data and the plurality of sets of classification data respectively produced by the plurality of previously generated machine learning models, thereby correcting the malicious shift of the updated current machine learning model,

generating a retrained current machine learning model, and

performing, by the computer, an action to prevent subsequent receipt of data associated with the entity so as to prevent a future input of data associated with the entity to at least the retrained current machine learning model, the action comprising at least one from among revoking the credential from the entity and blacklisting the entity.

2 . The method of claim 1 , wherein each of the plurality of previously generated machine learning models has an associated timestamp.

3 . The method of claim 1 , wherein the plurality of previously generated machine learning models are machine learning models created periodically over a period of one month or more.

4 . The method of claim 1 , further comprising:

retrieving, by the computer, the plurality of previously generated machine learning models from a model cache, wherein the plurality of previously generated machine learning models are selected from the model cache using a randomized sampling procedure.

5 . The method of claim 1 , wherein the trained current machine learning model and the plurality of previously generated machine learning models are associated with an entity profile.

6 . The method of claim 1 , wherein the trained current machine learning model is a neural network.

7 . A computer comprising:

a processor; and

a non-transitory computer readable medium coupled to the processor, the non-transitory computer readable medium comprising code that, when executed by the processor, causes the processor to perform a method including:

receiving a set of input data associated with an entity having a credential allowing the entity to communicate with the computer,

applying as an input, the set of input data to a trained current machine learning model,

in response to the input of the set of input data, producing, by the trained current machine learning model, a first set of classification data,

updating the trained current machine learning model using the set of input data and the first set of classification data to generate an updated current machine learning model,

determining first metadata corresponding to the trained current machine learning model before the updating, the first metadata corresponding to a metric or a characteristic of the trained current machine learning model before the updating,

determining second metadata corresponding to the updated current machine learning model, the second metadata corresponding to a metric or a characteristic of the updated current machine learning model,

analyzing the first metadata and the second metadata to determine a deviation metric,

determining that the deviation metric is greater than a deviation threshold,

in response to the deviation metric being greater than the deviation threshold:

applying the set of input data to a plurality of previously generated machine learning models, respectively, wherein the plurality of previously generated machine learning models do not include the updated current machine learning model,

producing, by the plurality of previously generated machine learning models, a plurality of sets of classification data in response to respectively applying the set of input data to the plurality of previously generated machine learning models,

determining whether the entity associated with the set of input data is a malicious entity attempting to induce a malicious shift of the trained current machine learning model toward incorrect classifications,

the determining including:

combining the plurality of sets of classification data to produce a combined set of classification data, wherein both the combined set of classification data and the first classification set of data are produced for a same set of input data,

analyzing the combined set of classification data and the first set of classification data to determine an error metric, wherein the error metric indicates a difference in (1) one or more classifications included in the first set of classification data produced by the updated current machine learning model and (2) classifications included in the combined set of classification data from the plurality of previously generated machine learning models, wherein the error metric is a measure of a classification divergence between the combined set of classification data and the first classification set of data,

determining that the error metric is greater than an error threshold, and

in response to the error metric being greater than the error threshold, determining that the entity associated with the set of input data is the malicious entity that is inducing the malicious shift of the trained current machine learning model toward the incorrect classifications, and

responsive to the determining that the entity is the malicious entity:

retraining the updated current machine learning model by providing, as an input, the set of input data and the plurality of sets of classification data respectively produced by the plurality of previously generated machine learning models, thereby correcting the malicious shift of the updated current machine learning model,

generating a retrained current machine learning model, and

performing an action to prevent subsequent receipt of data associated with the entity so as to prevent a future input of data associated with the entity to at least the retrained current machine learning model, the action including at least one from among revoking the credential from the entity and blacklisting the entity.

8 . The computer of claim 7 , wherein each of the plurality of previously generated machine learning models has an associated timestamp.

9 . The computer of claim 7 , wherein the plurality of previously generated machine learning models are machine learning models created periodically over a period of one month or more.

10 . The computer of claim 7 , wherein the method further includes:

retrieving, by the computer, the plurality of previously generated machine learning models from a model cache, wherein the plurality of previously generated machine learning models are selected from the model cache using a randomized sampling procedure.

11 . The computer of claim 7 , wherein the trained current machine learning model and the plurality of previously generated machine learning models are associated with an entity profile.