RU2809595C1 - Method and system for automatic polygraph testing using three ensembles of machine learning models - Google Patents

Abstract

FIELD: polygraph testing.

SUBSTANCE: invention relates to a method and system for automatically checking a subject using a polygraph using machine learning methods. In the method, records of polygraphic tests are obtained containing sensor signals with time scales on which the beginning and end of the question are marked; additional data containing at least the age of the person being checked, gender, job information are received; processing of the received signals using the first ensemble of ML models trained on one topic is performed, and during this processing the following is carried out: signal processing by the first ML model, during which the following is performed: determination of time intervals for extracting variables based on the time stamps of the beginning and end of the question and the time answer labels, and based on question type and topic; extracting variables from each signal at certain time intervals; processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them; feeding said vector to the 1st ML model to obtain the output value of the 1st ML model; transferring the output value of the 1st ML model to the input of the 2nd ML model; using the second ML model, the output value of the 1st ML model and additional data are processed, and during this processing the following is carried out: division of additional data into categorical and numerical variables; processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables; concatenation of processed additional variables, as well as the output value of the 1st ML model and construction of a vector based on them; feeding said vector to the 2nd ML model to obtain the output value of the 2nd ML model; feeding the output value of the 2nd ML model to the third ML model to form the output value of the first ensemble; carry out processing of the received signals using a second ensemble of the ML models trained on a combination of topics, and during this processing the following is carried out: signal processing by the first ML model during which the following is performed: determination of time intervals for extracting variables based on the time stamps of the beginning and end of the question and the time stamp answer, and based on the type and topic of the question; extracting variables from each signal at certain time intervals; processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them; feeding said vector to the 1st ML model to obtain the output value of the 1st ML model; transferring the output value of the 1st ML model to the input of the 2nd ML model; using the second ML model, the output value of the 1st ML model and additional data are processed, and during this processing the following is carried out: division of additional data into categorical and numerical variables; processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables; concatenation of processed additional variables, as well as the output value of the 1st ML model and construction of a vector based on them; feeding said vector to the 2nd ML model to obtain the output value of the 2nd ML model; feeding the output value of the 2nd ML model to the third ML model to form the output value of the second ensemble; carry out processing of the received signals using a third ensemble of machine learning models trained on a combination of topics, and during the specified processing the following is carried out: signal processing by the first ML model, during which the following is performed: determination of time intervals for extracting variables based on the time stamps of the beginning and end of the question and response timestamp, and based on question type and topic; extracting variables from each signal at certain time intervals; processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them; feeding said vector to the 1st ML model to obtain the output value of the 1st ML model; transferring the output value of the 1st ML model to the input of the 2nd ML model; using the second ML model, the output value of the 1st ML model and additional data are processed, and during this processing the following is carried out: division of additional data into categorical and numerical variables; processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables; concatenation of processed additional variables, as well as the output value of the 1st ML model, and construction of a vector based on them; feeding said vector to the 2nd ML model to obtain the output value of the 2nd ML model; feeding the output value of the 2nd ML model to the third ML model to form the output value of the third ensemble; using the third ML model, the output values of the first, second and third ML ensembles are processed, and during this processing the following is carried out: concatenation of the processed outputs data values of the first, second and third ensembles and construction of a vector based on them; feeding said vector to the 3rd ML model to obtain the output value of the 3rd ML model; comparison of the output value of the 3rd model with a given threshold value; and determining that the response is false if the output value is greater than or equal to the threshold value, or the response is true if the output value is below the threshold value.

EFFECT: increased accuracy of polygraph testing.

7 cl, 11 dwg, 11 tbl

Description

TECHNICAL FIELD

[0001] The claimed technical solution relates to an automated method and system for automatic printing verification using machine learning algorithms.

BACKGROUND OF THE ART

[0002] Classic polygraph screenings are regularly used by significant enterprises such as banks, law enforcement agencies and federal authorities. The main concern of the scientific community is that these screens are prone to error. However, these errors can be a consequence not only of the method, but also of the person (the polygraph examiner).

[0003] The security of customer money and data (eg transactions) is built into the core of banking culture and reputation. As one of the tools to protect clients, only with the consent of candidates and employees, and in accordance with the law, the bank uses polygraphic screenings (PS). They are used when hiring candidates in risky areas to prevent hiring an unreliable person. To detect violations, employees in high-risk positions are regularly tested. The CoP includes the following topics: drugs, gambling addiction, insider trading, disclosure of confidential information, bribery, corruption, misappropriation of funds and fraud. The financial industry is not the only one using PS; other examples are such important sectors as aviation, industry and law enforcement agencies around the world [1,2].

[0004] A classic polygraph is a device that records cardiovascular activity (such as heart rate), thoracic and abdominal respiration, galvanic skin response (skin electrical activity or SEA), and shivering. The polygraph examiner asks questions to the subject, to which he receives “yes” or “no” answers. Reviews of the classical polygraph and the methodology for constructing questions are presented in [3, 4, 5].

[0005] Non-traditional lie detection studies use video and audio analysis [6] (including facial expression [7, 8], pupil response [9] and question-answer latency [10]), electromyography (EMG) [11], electroencephalography (EEG) [12], magnetic resonance imaging [13, 14] or written sequences (keystroke dynamics) [15] in addition to classical polygraphic data.

[0006] Some of these studies have even expanded into new areas, such as the iBorderCtrl lie detector being tested at European airports [16, 17] or VeriPol being used by Spanish police in insurance claims cases [18, 19]. The classic polygraph remains the tool of choice in traditional tasks such as employment screening and criminal or internal investigations.

[0007] The polygraph has a long history of criticism from scientists in the fields of psychology and law, as well as from society and government [1, 23]. The main concern is that this technique does not reliably detect lies and truth. And yet, “paradoxically, while Congress expresses deep concern about the effectiveness of this technology, EPRA allows the use of lie detectors in cases where the accuracy of the result is of paramount importance: national defense, security and legitimate ongoing investigations” [22].

[0008] Criticisms of this technique provide many reasons why polygraph screening may fail to detect lies or flag truths as lies. For example, “Printing tests do not evaluate deception, but situations that are structured to evoke and evaluate fear” [24].

[0009] A truthful junior manager may fear being called corrupt more than a cold-blooded senior manager fears being caught in a lie by a polygraph examiner. Another example of constructive criticism is a call for standardization of the polygraph screening procedure and polygraph examiner training [25]. Polygraph examiner errors can occur, for example, when the polygraph examiner is inexperienced, tired, distracted, or biased [26].

[0010] There is a simple solution to quality control: always carry out the test by another polygraph examiner, who will confirm or refute the conclusion of the previous polygraph examiner [27]. To review a polygraph report, another polygraph examiner needs to review the screening recording, which includes the polygram (a graphical representation of the sensor data associated with the polygraph examiner's questions and the test taker's answers), sometimes audio and video recordings, and compare his report with the original. This check takes at least half the time of screening. A standard screening lasts a minimum of two hours. Thus, re-checking costs both time and money. For this reason, Homeland Security offices conduct follow-up inspections rarely or not at all. Another reason why tests by a second polygraph examiner may not be effective: the second polygraph examiner may make the same mistake that the original polygraph examiner made.

[0011] A common disadvantage of existing solutions in this area is the presence of a human factor during polygraphic checking, which negatively affects the accuracy and speed of verification, as well as the lack of an automated re-checking process.

DISCLOSURE OF INVENTION

[0012] The claimed technical solution proposes a new approach to automatic polygraph checking using machine learning (ML) models.

[0013] The effectiveness of this solution is confirmed by a significant increase in the accuracy and speed of automatic printing testing.

[0014] Thus, the technical problem of accurate and high-speed automatic printing verification is solved.

[0015] The technical result achieved by solving this problem is to increase the accuracy of printing testing.

[0016] An additional technical result achieved by solving this problem is increasing the speed of printing testing.

[0017] Also, an additional technical result achieved when solving this problem is the automation of the printing verification process.

[0018] These technical results are achieved by implementing a computer-implemented method for automatic polygraph checking, performed using a computing system containing at least three ensembles of machine learning models, wherein the method performs the steps of:

- obtain records of polygraph tests containing at least sensor signals with time scales on which the beginning and end of the question are marked;

- receive additional data containing at least the age of the person being checked, gender, job information;

- process the received signals using the first ensemble of ML models trained on one topic, and during this processing the following is carried out:

signal processing by the first MO model during which the following is performed:

defining time intervals for retrieving variables based on question start and end timestamps and answer timestamps, and based on question type and topic;

extracting variables from each signal at certain time intervals;

processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them;

feeding said vector to the 1st MO model to obtain the output value of the 1st MO model;

transferring the output value of the 1st MO model to the input of the 2nd MO model;

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

separating additional data into categorical and numerical variables;

processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables;

concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them;

feeding said vector to the 2nd MO model to obtain the output value of the 2nd MO model;

feeding the output value of the 2nd MO model to the third MO model to form the output value of the first ensemble;

- process the received signals using a second ensemble of ML models trained on a combination of topics, and during this processing the following is carried out:

signal processing by the first MO model during which the following is performed:

defining time intervals for retrieving variables based on question start and end timestamps and answer timestamps, and based on question type and topic;

extracting variables from each signal at certain time intervals;

processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them;

feeding said vector to the 1st MO model to obtain the output value of the 1st MO model;

transferring the output value of the 1st MO model to the input of the 2nd MO model;

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

separating additional data into categorical and numerical variables;

processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables;

concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them;

feeding said vector to the 2nd MO model to obtain the output value of the 2nd MO model;

feeding the output value of the 2nd MO model to the third MO model to form the output value of the second ensemble;

- process the received signals using a third ensemble of machine learning models trained on a combination of topics, and during this processing the following is carried out:

signal processing by the first MO model, during which the following is performed:

defining time intervals for retrieving variables based on question start and end timestamps and answer timestamps, and based on question type and topic;

extracting variables from each signal at certain time intervals;

processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them;

feeding said vector to the 1st MO model to obtain the output value of the 1st MO model;

transferring the output value of the 1st MO model to the input of the 2nd MO model;

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

separating additional data into categorical and numerical variables;

processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables;

concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them;

feeding said vector to the 2nd MO model to obtain the output value of the 2nd MO model;

feeding the output value of the 2nd MO model to the third MO model to form the output value of the third ensemble;

- using the third MO model, the output values of the first, second and third MO ensembles are processed, and during this processing the following is carried out:

concatenation of the processed output values of the first, second and third ensembles, and construction of a vector based on them;

feeding said vector to the 3rd MO model to obtain the output value of the 3rd MO model;

comparison of the output value of the 3rd model with a given threshold value; And

- determine that the response is false if the output value is greater than or equal to the threshold value or the response is true if the output value is below the threshold value.

[0019] In one of the private embodiments of the method, ML models are trained on one of the topics for checks or a combination thereof, where the topics for checks are: narcotic substances, receiving additional remuneration, disclosure of confidential information, debt obligations, third-party income, criminal offenses, administrative offenses, violations of internal regulations (INR).

[0020] In another particular embodiment of the method, the ML models of the first second and third ensembles are of a type selected from the group: gradient boosting, random forest, or neural network.

[0021] In another particular embodiment of the method, the third ML model has a type selected from the group: logistic regression, random forest, gradient boosting or averaging.

[0022] In another particular embodiment of the method for recording polygraphic checks, they contain signals from sensors, including at least one of: heart rate (HR), galvanic skin response (GSR), blood pressure, upper and lower respiration, piezoplethysmogram, photoplethysmogram , thermal, pupil movement or combinations thereof.

[0023] In another particular embodiment of the method, the additional data contains at least one of: identification number of polygraph examiners, identification number of polygraphs, information about weather conditions, results of an electroencephalogram, magnetic resonance imaging, functional near-infrared spectroscopy, information about geomagnetic storms or their combinations

[0024] In addition, the claimed technical result is achieved through an automatic printing verification system containing:

- at least one processor;

- at least one memory connected to the processor, which contains machine-readable instructions that, when executed by at least one processor, enable the execution of an automatic polygraphic checking method.

BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Features and advantages of the present invention will become apparent from the following detailed description of the invention and the accompanying drawings.

[0026] FIG. 1 illustrates an example of implementing a method for automatic polygraph checking using one ML model.

[0027] FIG. 2 illustrates an example of the implementation of a method for automatic polygraph checking using two ML models.

[0028] FIG. Figure 3 illustrates an example of the implementation of a method for automatic polygraph checking using two ensembles of ML models.

[0029] FIG. 4 illustrates an example of the implementation of a method for automatic polygraph checking using three ensembles of ML models.

[0030] FIG. Figure 5 illustrates a graph where the model and the polygraph examiner agreed and disagreed depending on the model’s assessment

[0031] FIG. 6 illustrates the distribution of ratings for two relevant topics.

[0032] FIG. 7 illustrates an example implementation of the basic model.

[0033] FIG. Figure 8 illustrates an example implementation of a model trained on one topic.

[0034] FIG. 9 illustrates an example of the implementation of a universal model.

[0035] FIG. 10 illustrates an example of signal convolution.

[0036] FIG. 11 illustrates a general view of the automatic printing verification system.

IMPLEMENTATION OF THE INVENTION

[0037] This technical solution can be implemented on a computer, in the form of an automated information system (AIS) or a machine-readable medium containing instructions for performing the above method.

[0038] The technical solution can be implemented in the form of a distributed computer system.

[0039] In this solution, a system means a computer system, a computer (computer), CNC (computer numerical control), PLC (programmable logic controller), computerized control systems and any other devices capable of performing a given, well-defined sequence of computing operations (actions, instructions).

[0040] A command processing device means an electronic unit or an integrated circuit (microprocessor) that executes machine instructions (programs)/

[0041] A command processing device reads and executes machine instructions (programs) from one or more storage devices, such as devices such as random access memory (RAM) and/or read only memory (ROM). ROM can be, but is not limited to, hard drives (HDD), flash memory, solid-state drives (SSD), optical storage media (CD, DVD, BD, MD, etc.), etc.

[0042] Program - a sequence of instructions intended for execution by a computer control device or command processing device.

[0043] For this technical solution, a basic second opinion model was built, trained on historical data from real answers from 2094 polygraph screenings, including the “False Detected” field filled in by the polygraph examiners who conducted the screening.

[0044] This technical solution demonstrates the quality metrics of the base model in Table 1.a, in the “all topics” column. The main metrics are ROC AUC (receiver operating characteristic) and TPR (target rating point) at FPR=0.05.

[0045] Tables 2 and 3 reflect the importance of features based on 10 physiological signals and features of gender and age of the subject; details of the construction of second-level features will be described below. Fig. Figure 5 displays a graph where the model and polygraph examiner agreed and disagreed depending on the model's assessment.

[0046] The quality of the model was also measured by applying it to each of the seven screening topics (Table 1.a).

[0047] Evaluating the performance of the model entails the use of approximately 100 man-hours because it would require the use of an actual polygraph examiner who would review multiple polygraph screenings selected based on the evaluation of the model.

[0048] It has also been suggested that information about geomagnetic storms on Earth [28] and weather conditions in the city on the day of screening could improve the model's prediction quality. The basis for this assumption was that during geomagnetic storms and under different weather conditions, people may behave differently, which may result in a slight change in physiological signals, or the sensor data may be slightly biased, or both.

[0049] The ID of the polygraph examiner was also obtained in the belief that this data may help the model, since different polygraph examiners may produce slightly different physiological responses in subjects, or the printing equipment assigned to each polygraph examiner may have different measurement errors. Data was also collected on the subjects' job titles because people with different education and training may tell the truth and lie differently.

[0050] Table l.b shows the quality of the base model trained with alternative data, and Table 4 shows the importance of these features. The improvements from each source relative to the model built only on physiological signal data are shown in Table 5.

[0051] As the experiment shows, all alternative data improve the quality of the model, but only age, gender, position and geomagnetic storms were retained for operation (Table 1.c).

[0052] Weather shows a high increase in quality and high importance, and there was concern that this might be due to technical reasons: the data has a strong imbalance in the proportion of detected by city. To eliminate bias caused by city imbalance, data was taken from only one city, but weather still had a high importance value.

[0053] Thus, it can be concluded that weather is a meaningful alternative data. However, across the entire sample, weather may carry information about the city, and the model estimate may be biased due to an imbalance of those identified across cities. While the polygraph examiner ID feature shows a moderate increase in quality, the nature of this feature requires deeper study before using this feature in the operation of the model.

[0054] For example, if in fact it is not the polygraph examiner ID that contributes to quality improvement, but the polygraph ID, then when one polygraph examiner changes the device, an incorrect assessment of the model will be obtained.

[0056] If you train a separate model for each topic, the results will be better than the results of the basic model with geomagnetic storms. Table 6 shows that +2% ROC AUC (6% relative gain) was obtained if the model was trained only on the drug topic. Thus, dividing into topics will help the model to learn and make a conclusion better.

[0058] The drug-trained model discussed in the previous paragraph was applied to all other six topics (Table 6). Compared to the universal model, the quality of the drug model varies from topic to topic. This proves that a model trained on one topic can be used on other topics, albeit with minor drops in quality for some topics.

[0059] Table 1 shows that the model performs better on some topics (such as drugs and criminal offenses) than on other topics (such as outside income and GNI violations). This observation is consistent with a long-standing problem with screenings: people cannot answer a question with confidence when they are not completely sure of the answer.

[0060] For example, there are hundreds of IRRs in a bank, each with many pages, and that's without mentioning the different versions, so people may not be sure whether they violated at least one regulatory document. A similar situation is observed with third-party income: some people begin to ask questions like “if I received money from relatives, is this third-party income?” and so on.

[0061] This observation can also be used to improve the quality of the base model. If topics are found that people are unsure about, then the model should be unsure by learning from those topics.

[0062] Vague topics were removed from the entire sample, however, Table 7 shows that this action did not significantly improve the quality of the model, and the quality of scoring “vague” topics dropped or remained unchanged. This is due to the fact that the proportion of “vague” topics identified in screenings in the training sample is insignificant.

[0063] The constructed universal model, which was described above, does not use the topic name as a feature for training and evaluation. The reason for this is that a model was needed that can evaluate any topic in the screening, not just the 7 topics that are present in the training set.

[0064] Table 11 below shows how using the topic name as an additional feature affects the quality of the model. It is shown that knowledge of the topic helps the model evaluate the “blurry” topic of violation of GNI, while the quality on other topics remains unchanged.

[0065] Before adding the “topic name” attribute, the sample was balanced by the number of identified risk factors in screenings by topic. This balancing consists of reducing the number of narcotic substances identified on the topic by 4 times. Table 11 shows that this reduction reduces the quality of the drug theme, which previously was always the unexplained leader.

[0066] The quality on the topic of narcotic substances was equal to the topics closest in quality, such as receiving monetary rewards and criminal and administrative offenses. This observation explains the previous leadership of the NV topic - this was due to the fact that this topic had the largest number of objects of a minor class (hiding of information was revealed) compared to other topics.

[0067] The improvement in quality from adding new, additional screens with identified risk factors was measured, and experiments were conducted with various variations of the ensembles. The results are presented in Table 8.

[0068] Combined, ensemble and additional data increased AUC by 5% across all topics and by 11% on selected topics.

[0069] Now you can compare the performance of two advanced models: the Universal model (ensemble with alternative data) and the model trained on the NV topic (model on the same topic with alternative data). Testing of polygraph examiners' error detection is carried out among 2094 archival screenings. Polygraph examiners search for errors in two topics that are most common in screenings. Screenings were selected in which the polygraph examiner concluded that the topic was not identified, but the model confidently stated that a lie was identified.

[0070] Based on the 15 highest scores of the model trained on the Substances topic, 15 polygraph examiner findings with a “Not Detected” conclusion were selected as candidates for polygraph examiner error on the Narcotics Substances topic. In the same way, based on the estimates of the universal model, 15/5/5 conclusions “Not identified” were selected for the topics “Receipt of monetary reward” / “Disclosure of confidential information” / “Criminal and administrative offenses”. Thus, 40 conclusions were obtained from 36 screenings.

[0071] Data from 36 screenings were verified through blind retesting by two polygraph examiners. The polygraph examiners did not know the results of the screenings and did not share the results with each other. The reason for conducting two tests is that in the event of a discrepancy between the conclusions of the original report and one of the repeated tests, it will be the word of one polygraph examiner against the other.

[0072] The experiment involved extremely experienced polygraph examiners, and it is believed that the error rate of polygraph examiners among all screeners ranges from 0.0 to 1.0%.

[0073] An overview of the two replicate tests is presented in Table 9. The distributions of scores for the two relevant topics are presented in FIG. 6.

[0074] By reviewing 39 findings in 35 screenings (out of 2094 screenings), 30 problematic findings were identified in which the polygraph examiner's error was confirmed by two tests (13 findings) or one of the tests did not agree with the original conclusion (17 findings). The remaining 9 conclusions are model errors, i.e. the original conclusion was correct, which was confirmed by both repeated checks.

[0075] It was expected that there would be cases of discrepancy between the results of two repeated tests, because There are difficult cases in which the solution is not obvious. For such difficult cases, a consultation is held at which polygraph examiners discuss their conflicting conclusions and come to a consensus.

[0076] It has been observed that in some problematic screenings (finding “Revealed” in one or both retests), the polygraph examiners who performed the retest noted that the examinee was resisting. Omitting a countermeasure is, by definition, a polygraph examiner's error.

Description of the data.

[0077] The dataset consists of 2094 historical polygraph tests of the screening type, including the labels “Concealment of information detected” and “Concealment of information not detected”, put down by polygraph examiners who conducted the screenings. These screenings were carried out on candidates and bank staff in risky areas, with their consent and in accordance with the law, before hiring, promotion or every year, depending on the position. The polygraph screening includes a subset of 14 topics, including drugs and receipt of monetary reward.

[0078] The screening recording includes the subject's physiological signals, audio, and questions in text format. Each question has three time stamps for each presentation: the beginning of the question asked by the polygraph examiner, the end of the question, and the moment of the answer. Each question is classified into four types.

[0079] Question types:

[0080] Examples of questions:

[0081] In addition to the physiological signals, Table 10 records the gender and age of the subject.

[0082] Screening was carried out on a Polyconius polygraph, model 7.

Data processing

[0083] The main task is to evaluate the polygraph examiner's conclusion (Detected or Not Detected) for a certain topic in the screening.

[0084] To build the model, data was presented in the following format: each row in the data set represents a screening record on a specific topic on a specific test for one examinee. The target variable is labeled as “Detected” or “Not Detected”. There may be a bias due to such marking, since the test taker may not lie on all texts on the topic during screening.

[0085] Physiological signals were extracted from time windows according to question timestamps. Thus, initially a line is a time series of a physiological signal for a specific presentation on a specific topic.

[0086] For each presentation of the test and control questions, basic statistics were generated: minimum, maximum, mean, amplitude and standard deviation. Next, the minimum, maximum, mean and standard deviation were used as aggregate functions at each step. Data from each presentation were grouped by question. An additional feature was created that characterizes the difference between the first and subsequent presentations. Similarly, questions were grouped into topics within each test. Ultimately, each line includes 600 features extracted from the polygraph screening record for a specific topic in a specific test (Figure 10).

Models.

[0087] In this technical solution, a stacking ensemble was used, which consisted of two levels of models based on the gradient boosting algorithm. The two-level structure was required in order to avoid the “curse of dimensionality.” The first level model was trained on 600 features built on physiological signals for each topic, making a Detected/Not Detected conclusion for each test within the screen. The output values of the first model aggregated on each topic were fed to the input of the second model.

[0088] The second level model had the following features:

pred_proba_max - maximum probability of detecting lies among tests;

pred_proba_mean - average probability of detecting a lie;

pred_proba_min - the minimum probability of detecting lies in the test environment;

pred_proba_diff - the difference between the maximum and average probabilities.

[0089] These probabilities are combined with alternative data (biographical data, data on magnetic storms, etc.) The resulting data set is fed into the input of a second model, which produces the probability of lying during screening on a specific topic.

Basic model.

[0090] This model (Figure 7) does not receive topic information during training and prediction. The topic information is used for further aggregation within the screening. For example, in the first screening, aggregation occurs on the topic “drugs” across all tests, the same thing happens, for example, with the topic “insider trading” in the second screening.

Model trained on one topic

[0091] The logic for building the model and creating features is the same as in the base model. The difference is that for training, features are taken for only one topic. Data from only one topic are selected before the first level model is applied. In FIG. Figure 8 shows the order of training the model for the topic narcotic substances, while the remaining topics were removed from the training data set.

Universal model.

[0092] It was decided to use the best features of the models described above. Thus, an ensemble of existing architectures was built (Fig. 9). After a series of experiments, the best result was shown by an architecture that averaged the degree of confidence of the following models:

basic model - algorithm: gradient boosting; with alternative data; model on one topic - algorithm: gradient boosting; with alternative data; basic model - algorithm: random forest.

[0093] This ensemble was applied to all topics except the Drugs topic. In addition to the usual advantages of ensembles, it was rational to use different architectures (gradient boosting and random forest) in order to exclude pure model error, and those screenings in which there was a polygraph examiner error would be highlighted.

[0094] As shown in FIG. 1 method of automatic polygraph checking (100), is performed using a computing system containing a machine learning model, and consists of several interrelated stages.

[0095] At step 101, polygraph test records are obtained containing at least sensor signals with time scales on which the beginning and end of the question are marked.

[0096] Next, at step 102, additional data is obtained containing at least the age of the person being checked, gender, job information.

[0097] Next, at step 103, the received signals and additional data are processed using a machine learning (ML) model, and during the said processing:

defining time intervals for retrieving variables based on question start and end timestamps and answer timestamps, and based on question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected;

extracting variables from each signal at specific time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database;

processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

separating additional data into categorical and numerical variables.

processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

concatenating the processed variables extracted from each signal and the processed additional variables, and constructing a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected);

feeding said vector into the ML model to obtain the output value of the ML model. At this stage, model speeds are obtained in the range from 0 to 1;

comparing the output value of the model with a specified threshold value.

[0098] And at step 104, it is determined that the response is false if the output value is greater than or equal to the threshold value or the response is true if the output value is below the threshold value.

[0099] In one of the particular embodiments of the method, the MO model determines the answer as true or false not by one received answer, but by a set of questions and answers related to the topic of polygraph testing, while the received variable signals are combined into a single vector, or averaged before transfer to the MO model.

[0100] In another particular embodiment of the method, the ML model is of the gradient boosting, random forest, or neural network type.

[0101] In another particular embodiment of the method, the ML model is trained on one of the topics for checks or a combination thereof, where the topics for checks are: narcotic substances, receiving additional remuneration, disclosure of confidential information, debt obligations, third-party income, criminal offenses, administrative offenses , violations of internal regulatory documents (INR).

[0102] In another particular embodiment of the method for recording polygraphic checks, they contain signals from sensors, including at least one of: heart rate (HR), galvanic skin response (GSR), blood pressure, upper and lower respiration, piezoplethysmogram, photoplethysmogram , thermal, pupil movement or combinations thereof.

[0103] In another particular embodiment of the method, the additional data contains at least one of: identification number of polygraph examiners, identification number of polygraphs, information about weather conditions, results of an electroencephalogram, magnetic resonance imaging, functional near-infrared spectroscopy, information about geomagnetic storms or their combinations.

[0104] The implementation of this technical solution based on one MO model allows you to automate the process of polygraphic checks and detect information concealment with high accuracy.

[0105] As shown in FIG. 2, in another particular embodiment, the automatic polygraph checking method (200) is performed using a computing system containing at least two machine learning models and consists of several interrelated steps.

[0106] At step 201, polygraph test records are obtained containing at least sensor signals with time scales on which the beginning and end of the question are marked.

[0107] Next, at step 202, additional data is obtained containing at least the age of the person being checked, gender, and job information.

[0108] Next, at step 203, the received signals are processed using the first machine learning (ML) model, and during the said processing:

defining time intervals for retrieving variables based on question start and end timestamps and answer timestamps, and based on question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected;

extracting variables from each signal at specific time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database;

processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

feeding the said vector to the 1st MO model to obtain the output value of the 1st MO model. At this stage, model speeds are obtained in the range from 0 to 1;

transferring the output value of the 1st MO model to the input of the 2nd MO model.

[0109] Next, at step 204, using a second machine learning (ML) model, the output value of the 1st ML model and additional data are processed, and during the specified processing:

separating additional data into categorical and numerical variables;

processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected);

feeding the vector to the 2nd MO model to obtain the output value of the 2nd MO model. At this stage, model speeds are obtained in the range from 0 to 1;

comparison of the output value of the 2nd model with a given threshold value.

[0110] And at step 205, it is determined that the response is false if the output value is greater than or equal to the threshold value or the response is true if the output value is below the threshold value.

[0111] In one of the particular embodiments of the method, the ML model determines the answer as true or false not by one received answer, but by a set of questions and answers related to the topic of polygraph testing, while the received variable signals are combined into a single vector, or averaged before transfer to the MO model.

[0112] In another particular embodiment of the method, the 1st ML model is of the type gradient boosting, random forest, or neural network.

[0113] In another particular embodiment of the method, the 2nd ML model is of the type gradient boosting, random forest, or neural network.

[0114] In another particular embodiment of the method, ML models are trained on one of the topics for checks or a combination thereof, where the topics for checks are: narcotic substances, receiving additional remuneration, disclosure of confidential information, debt obligations, third-party income, criminal offenses, administrative offenses , violations of internal regulatory documents (INR).

[0115] In another particular embodiment of the method for recording polygraphic checks, they contain signals from sensors, including at least one of: heart rate (HR), galvanic skin response (GSR), blood pressure, upper and lower respiration, piezoplethysmogram, photoplethysmogram , thermal, pupil movement or combinations thereof.

[0116] In another particular embodiment of the method, the additional data contains at least one of: identification number of polygraph examiners, identification number of polygraphs, information about weather conditions, results of an electroencephalogram, magnetic resonance imaging, functional near-infrared spectroscopy, information about geomagnetic storms or their combinations.

[0117] The implementation of this technical solution based on two MO models allows you to automate the process of polygraphic checks and identify hidden information with high accuracy. This implementation makes it possible to increase the accuracy of identifying information hiding due to the preliminary calculation of physiological data in the first level model and the use of aggregates from the first model (with the addition of additional non-physiological data) in the 2nd level model.

[0118] As shown in FIG. 3, in another particular embodiment, the automatic polygraph verification method (300) is performed using a computing system containing at least two ensembles of machine learning models, and consists of several interrelated stages.

[0119] At step 301, polygraph test records are obtained containing at least sensor signals with time lines on which the beginning and end of the question are marked.

[0120] Next, at step 302, additional data is obtained containing at least the age of the person being checked, gender, and job information.

[0121] Next, at step 303, the received signals are processed using the first ensemble of ML models trained on one topic, and during this processing the following is carried out:

signal processing by the first MO model during which the following is performed:

• defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp, and based on the question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected;

• extracting variables from each signal at certain time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database;

• processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

• feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model. At this stage, model speeds are obtained in the range from 0 to 1;

• transfer of the output value of the 1st MO model to the input of the 2nd MO model.

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

• dividing additional data into categorical and numerical variables;

• processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

• concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected);

• feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model. At this stage, model speeds are obtained in the range from 0 to 1;

• feeding the output value of the 2nd MO model to the third MO model to form the output value of the first ensemble.

[0122] Next, at step 304, the received signals are processed using a second ensemble of ML models trained on a combination of topics, and during the specified processing:

signal processing by the first MO model during which the following is performed:

• defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp, and based on the question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected;

• extracting variables from each signal at certain time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database;

• processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

• feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model. At this stage, model speeds are obtained in the range from 0 to 1;

• transfer of the output value of the 1st MO model to the input of the 2nd MO model.

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

• dividing additional data into categorical and numerical variables;

• processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1;

• concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected);

• feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model. At this stage, model speeds are obtained in the range from 0 to 1;

• supplying the output value of the 2nd MO model to the third MO model to form the output value of the second ensemble.

[0123] Next, at step 305, using the third MO model, the output values of the first and second MO ensembles are processed, and during this processing the following is carried out:

concatenation of the processed output values of the first and second ensembles, and construction of a vector based on them.

feeding the said vector into the 3rd MO model to obtain the output value of the 3rd MO model. At this stage, model speeds are obtained in the range from 0 to 1;

comparison of the output value of the 3rd model with a given threshold value.

[0124] And at step 306, it is determined that the response is false if the output value is greater than or equal to the threshold value or the response is true if the output value is below the threshold value.

[0125] In one of the private embodiments of the method, ML models are trained on one of the topics for checks or a combination thereof, where the topics for checks are: narcotic substances, receiving additional remuneration, disclosure of confidential information, debt obligations, third-party income, criminal offenses, administrative offenses, violations of internal regulations (INR).

[0126] In another particular embodiment of the method, the ML models of the first and second ensembles are of a type selected from the group: gradient boosting, random forest, or neural network.

[0127] In another particular embodiment of the method, the third ML model is of a type selected from the group: logistic regression, random forest, gradient boosting or averaging.

[0128] In another particular embodiment of the method for recording polygraphic checks, they contain signals from sensors, including at least one of: heart rate (HR), galvanic skin response (GSR), blood pressure, upper and lower respiration, piezoplethysmogram, photoplethysmogram , thermal, pupil movement or combinations thereof.

[0129] In another particular embodiment of the method, the additional data contains at least one of: identification number of polygraph examiners, identification number of polygraphs, information about weather conditions, results of an electroencephalogram, magnetic resonance imaging, functional near-infrared spectroscopy, information about geomagnetic storms or their combinations.

[0130] The implementation of this technical solution based on two ensembles of MO models allows you to automate the process of polygraphic checks and identify information concealment with high accuracy. This implementation makes it possible to increase the accuracy of detecting information hiding due to differences in architectures between the two ensembles and by adding a top-level model.

[0131] As shown in FIG. 3, in another particular embodiment, the automatic polygraph verification method (400) is performed using a computing system containing at least three ensembles of machine learning models, and consists of several interrelated stages.

[0132] At step 401, polygraph test records are obtained containing at least sensor signals with time scales marking the beginning and end of the question.

[0133] Next, at step 402, additional data is obtained containing at least the age of the person being checked, gender, job information.

[0134] Next, at step 403, the received signals are processed using the first ensemble of ML models trained on one topic, and during this processing the following is carried out:

signal processing by the first MO model, during which the following is performed:

• defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp, and based on the question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected.

• extracting variables from each signal at certain time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database.

• processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1.

• feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model. At this stage, model speeds in the range from 0 to 1 are obtained.

• transfer of the output value of the 1st MO model to the input of the 2nd MO model.

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

• separating additional data into categorical and numerical variables.

• processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1.

• concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected).

• feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model. At this stage, model speeds in the range from 0 to 1 are obtained.

• feeding the output value of the 2nd MO model to the third MO model to form the output value of the first ensemble.

[0135] Next, at step 404, the received signals are processed using a second ensemble of ML models trained on a combination of topics, and during the specified processing:

signal processing by the first MO model during which the following is performed:

• defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp, and based on the question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected.

• extracting variables from each signal at certain time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database.

• processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of the numerical variable becomes equal to a value from 0 to 1. Feeding the said vector into the 1st ML model to obtain the output value of the 1st ML model. At this stage, model speeds in the range from 0 to 1 are obtained.

• transfer of the output value of the 1st MO model to the input of the 2nd MO model.

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

• separating additional data into categorical and numerical variables.

• processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1.

• concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected).

• feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model. At this stage, model speeds in the range from 0 to 1 are obtained.

• supplying the output value of the 2nd MO model to the third MO model to form the output value of the second ensemble.

[0136] Next, at step 405, the received signals are processed using a third ensemble of machine learning models trained on a combination of topics, and during the said processing:

signal processing by the first MO model, during which the following is performed:

• defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp, and based on the question type and topic. At this stage, depending on the logic of building the model, time stamps are allocated only for those questions (and answers) whose physiological data will be involved in training and testing the model. For example, the generic model uses questions on all topics, so all labels will be selected.

• extracting variables from each signal at certain time intervals. At this stage, physiological signals at previously selected intervals are extracted from the database.

• processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1.

• feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model. At this stage, model speeds in the range from 0 to 1 are obtained.

• transfer of the output value of the 1st MO model to the input of the 2nd MO model.

using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out:

• separating additional data into categorical and numerical variables.

• processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1.

• concatenation of processed additional variables, as well as the output value of the 1st MO model, and construction of a vector based on them. At this stage, from all the selected data relevant to a specific polygraph examiner’s conclusion, one vector is obtained, which is equally applicable both for training the model (if the polygraph examiner’s conclusion is available) and for making a decision by the model (detected / not detected).

• feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model. At this stage, model speeds in the range from 0 to 1 are obtained.

• supplying the output value of the 2nd MO model to the third MO model to form the output value of the third ensemble.

[0137] Next, at step 406, using the third MO model, the output values of the first, second and third MO ensembles are processed, and during this processing the following is carried out:

concatenation of the processed output values of the first, second and third ensembles, and construction of a vector based on them. At this stage, each value of a numerical variable becomes equal to a value between 0 and 1.

feeding the said vector into the 3rd MO model to obtain the output value of the 3rd MO model. At this stage, model speeds in the range from 0 to 1 are obtained.

comparison of the output value of the 3rd model with a given threshold value.

[0138] And at step 407, a determination is made that the response is false if the output value is greater than or equal to the threshold value or the response is true if the output value is below the threshold value.

[0139] In one of the private embodiments of the method, ML models are trained on one of the topics for checks or a combination thereof, where the topics for checks are: narcotic substances, receiving additional remuneration, disclosure of confidential information, debt obligations, third-party income, criminal offenses, administrative offenses, violations of internal regulations (INR).

[0140] In another particular embodiment of the method, the ML models of the first, second and third ensembles are of a type selected from the group: gradient boosting, random forest, or neural network.

[0141] In another particular embodiment of the method, the third ML model is of a type selected from the group: logistic regression, random forest, gradient boosting, or averaging.

[0142] In another particular embodiment of the method for recording polygraphic checks, they contain signals from sensors, including at least one of: heart rate (HR), galvanic skin response (GSR), blood pressure, upper and lower respiration, piezoplethysmogram, photoplethysmogram , thermal, pupil movement or combinations thereof.

[0143] In another particular embodiment of the method, the additional data contains at least one of: identification number of polygraph examiners, identification number of polygraphs, information about weather conditions, results of an electroencephalogram, magnetic resonance imaging, functional near-infrared spectroscopy, information about geomagnetic storms or their combinations.

[0144] The implementation of this technical solution based on three ensembles of MO models allows you to automate the process of polygraphic checks and identify information concealment with high accuracy. This implementation makes it possible to increase the accuracy of detecting information hiding due to differences in architectures between the three ensembles and by adding a top-level model.

[0145] In FIG. 11 shows an example of a general view of a computing system (500), which implements the claimed method or is part of a computer system, for example, a server, a personal computer, or part of a computing cluster that processes the necessary data to implement the claimed technical solution.

[0146] In general, the system (500) contains one or more processors (501), memory devices such as RAM (302) and ROM (503), input/output interfaces (504), and input devices connected by a common information exchange bus. /output (1105), and a device for network communication (506).

[0147] The processor (501) (or multiple processors, multi-core processor, etc.) may be selected from a variety of devices commonly used today, for example, from manufacturers such as: Intel™, AMD™, Apple™, Samsung Exynos ™, MediaTEK™, Qualcomm Snapdragon™, etc. The processor or one of the processors used in the system (500) must also include a graphics processor, such as an NVIDIA or Graphcore GPU, the type of which is also suitable for full or partial implementation of the method, and can also be used to train and apply machine learning models in various information systems.

[0148] RAM (502) is a random access memory and is designed to store computer-readable instructions executable by the processor (501) to perform the necessary logical data processing operations. RAM (502) typically contains executable operating system instructions and related software components (applications, program modules, etc.). In this case, the available memory capacity of the graphics card or graphics processor can act as RAM (502).

[0149] A ROM (503) is one or more permanent storage devices, such as a hard disk drive (HDD), a solid state drive (SSD), flash memory (EEPROM, NAND, etc.), optical storage media ( CD-R/RW, DVD-R/RW, BlueRay Disc, MD), etc.

[0150] To organize the operation of system components (500) and organize the operation of external connected devices, various types of I/O interfaces (504) are used. The choice of appropriate interfaces depends on the specific design of the computing device, which can be, but is not limited to: PCI, AGP, PS/2, IrDa, Fire Wire, LPT, COM, SATA, IDE, Lightning, USB (2.0, 3.0, 3.1, micro , mini, type C), TRS/Audio jack (2.5, 3.5, 6.35), HDMI, DVI, VGA, Display Port, RJ45, RS232, etc.

[0151] To ensure user interaction with the computing system (300), various means (505) of I/O information are used, for example, a keyboard, a display (monitor), a touch display, a touch pad, a joystick, a mouse, a light pen, a stylus, touch panel, trackball, speakers, microphone, augmented reality tools, optical sensors, tablet, light indicators, projector, camera, biometric identification tools (retina scanner, fingerprint scanner, voice recognition module), etc.

[0152] The networking facility (506) enables data transmission via an internal or external computer network, such as an Intranet, Internet, LAN, or the like. One or more means (306) may be used, but not limited to: Ethernet card, GSM modem, GPRS modem, LTE modem, 5G modem, satellite communication module, NFC module, Bluetooth and/or BLE module, Wi-Fi module and etc.

[0153] The submitted application materials disclose preferred examples of implementation of a technical solution and should not be interpreted as limiting other, particular examples of its implementation that do not go beyond the scope of the requested legal protection, which are obvious to specialists in the relevant field of technology.

Information sources:

[1] M. Harris, "The Lie Generator: Inside the Black Mirror World of Polygraph Job Screenings," in Wired.com, 2018.

[2] B. Banerjee and G. Chatterjee, “The world of lie detection: a study into state of lie detection usage by state and society in Asia, Africa and Europe,” 2021.

[3] S. E. Fienberg, J. J. Blascovich, J. T. Cacioppo, R. J. Davidson, P. Ekman, D. L. Faigman and e. al., "The polygraph and lie detection," in National Research Council, The National Academies Press, Washington, DC, 2003.

[4] A. Slavkovic, "Evaluating polygraph data," in Carnegie Mellon University, 2002.

[5] J. Synnott, D. Dietzel and M. Ioannou, “The Polygraph: History, Methodology and Current Status,” in Reviewing Crime Psychology, 2020.

[6] G. Krishnamurthy, N. Majumder, S. Poria and E. Cambria, "A Deep Learning Approach for Multimodal Deception Detection," in 19th International Conference on Computational Linguistics and Intelligent Text Processing (CICLing), Hanoi, Vietnam, 2018 .

[7] D. Avola, L. Cinque, G. L. Foresti and D. Pannone, “Automatic Deception Detection in RGB videos using Facial Action Units,” in ICDSC 2019: Proceedings of the 13th International Conference on Distributed Smart Cameras, Trento, 2019.

[8] N. Samadiani, G. Huang, B. Cai, W. Luo, C.-H. Chi, Y. Xiang and J. He, "A Review on Automatic Facial Expression Recognition Systems Assisted by Multimodal Sensor Data," in Sensors, Vol.19, Issue 8, Sp.Issue Sensor Applications on Facial Analysis, 2019.

[9] A.K. Webb, C.R. Honts, J.C. Kircher, P. Bernhardt and A.E. Cook, "Effectiveness of pupil diameter in a probable-lie comparison question test for deception," in Legal and Criminological Psychology, Vol. 14, Issue 2, 2010.

[10] J.J. Walczyk, K.T. Mahoney, D. Doverspike and D.A. Griffith-Ross, "Cognitive Lie Detection: Response Time and Consistency of Answers as Cues to Deception," in Journal of Business and Psychology, Vol.24, 2009.

[11] A. Shuster, L. Inzelberg, O. Ossmy, L. Izakson, Y. Hanein and D. Levy, “Lie to my face: An electromyography approach to the study of deceptive behavior,” in Brain and Behavior, vol. . 11 issue 12, 2021.

[12] V. Abootalebi, M.H. Moradi and M.A. Khalilzadeh, "A new approach for EEG feature extraction in P300-based lie detection," in Computer Methods and Programs in Biomedicine, vol.94, issue 1, 2009.

[13] A. Kozel, K. Johnson, Q. Mu, E. Grenesko, S. Laken and M. George, "Detecting Deception Using Functional Magnetic Resonance Imaging," in Biological Psychiatry, Vol.58, Issue 8, 2005.

[14] M. J. Farah, J. B. Hutchinson, E. A. Phelps and A. D. Wagner, “Functional MRI-based lie detection: scientific and societal challenges,” in Nature Reviews Neuroscience, Vol. 14, 2014.

[15] M. Monaro, C. Galante, R. Spolaor, Q.Q. Li, L. Gamberini, M. Conti and G. Sartori, "Covert lie detection using keyboard dynamics," in Nature, Scientific Reports, Vol.8, 2018.

[16] L. Sousedikova, M. Hromada and M. Adamek, "Analysis of Artificial Intelligence Lie Detector Developed for Airport Security," in Tomas Bata University in Zlin, 2021.

[17] and L. Dencik, "The politics of deceptive borders: 'biomarkers of deceit and the case of iBorderCtrl," in Information, Communication & Society, Vol.25 Issue 3, 2022.

[18] F. Liberatore, J. Camacho-Collados and M. Camacho-Collados, "Applying automatic text-based detection of deceptive language to police reports: Extracting behavioral patterns from a multi-step classification model to understand how we lie to the police," in Knowledge-Based Systems, Vol.749,2018.

[19] "Police use a computer to expose false testimony. A lie-detection system being used by Spanish police highlights concerns about algorithms." in Nature, Editorial, 2018.

[20] G. Ben-Shakhar and W. Iacono, “Fallacies in the estimation of the validity of the Comparison Question Polygraph Test: A reply to Ginton (2020),” in Investigative Psychology and Offender Profiling, Vol.18, Issue 3 , 2021.

[21] D. Grubin and L. Madsen, “Lie detection and the polygraph: A historical review,” in The Journal oj Forensic Psychiatry & Psychology, 2005.

[22] C. Hinkle, “The Modern Lie Detector: AI-Powered Affect Screening and the Employee Polygraph Protection Act (EPPA),” in The Georgetown Law Journal Vol.109, Georgetown, 2021.

[23] J. Bittle, “Lie detectors have always been suspect. AI has made the problem worse.” in MIT Technology review, 2020.

[24] L. Saxe, "Science and the CQT polygraph - A theoretical critique," in Integrative Physiological and Behavioral Science, 1991.

[25] A. M. Perkey, “Recommendations for uniform polygraph examinations for preemployment screening of law enforcement applicants,” in University of Wisconsin-Platteville, 2021.

[26] W. Egerton, "Use of the Polygraph to Screen Police Candidates," in Law Enforcement Management Institute of Texas (LEMIT), North Richland Hills, Texas, 2020.

[27] D. Baur, "Federal Psychophysiological Detection of Deception Examiner Handbook," in Counterintelligence Field Activity Technical Manual, 2006.

[28] J. Matzka, O. Bronkalla, K. Tornow, K. Elger and C. Stolle, “Geomagnetic Kp index V. 1.0,” in GFZ Data Services. https://doi.org/10.5880/Kp.0001, Potsdam, Germany, 2021.

[29] C.R. Honts and S. Amato, "Automation of a screening polygraph test increases accuracy," in Psychology, Crime & Law, Vol.13, Issue 2, 2007.

[30] A. Mambreyan, E. Punskaya and H. Gunes, “Dataset Bias in Deception Detection,” in 26TH International Conference on Pattern Recognition, Montreal Quebec, 2022.

[31] M. Abouelenien, R. Mihalcea and M. Burzo, "Detecting Deceptive Behavior via Integration of Discriminative Features from Multiple Modalities," in IEEE Transactions on Information Forensics and Security, Vol. 12, Issue 5, 2016.

[32] Interfax, “Bill on possible ban on transfer abroad of Russians” personal data being submitted to State Duma,” in https://interfax.com/newsroom/top-stories/77833/, 2022.

[33] M. Handler and N. Hernandez, "Introduction to the NCCA ASCII Standard," in Polygraph & Forensic Credibility Assessment: A Journal of Science and Field Practice, 2019.

Claims (55)

1. A computer-implemented method for automatically testing a subject using a polygraph using machine learning methods, containing stages in which: - obtain records of polygraph tests containing at least sensor signals with time scales on which the beginning and end of the question are marked; - receive additional data containing at least the age of the person being checked, gender, job information; - process the received signals using the first ensemble of ML models trained on one topic, and during this processing the following is carried out: signal processing by the first MO model, during which the following is performed: • defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp and based on the question type and topic; • extracting variables from each signal at certain time intervals; • processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them; • feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model; • transfer of the output value of the 1st MO model to the input of the 2nd MO model; using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out: • dividing additional data into categorical and numerical variables; • processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables; • concatenation of processed additional variables, as well as the output value of the 1st MO model and construction of a vector based on them; • feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model; • feeding the output value of the 2nd MO model to the third MO model to form the output value of the first ensemble; - process the received signals using a second ensemble of ML models trained on a combination of topics, and during this processing the following is carried out: signal processing by the first MO model, during which the following is performed: • defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp and based on the question type and topic; • extracting variables from each signal at certain time intervals; • processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them; • feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model; • transfer of the output value of the 1st MO model to the input of the 2nd MO model; using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out: • dividing additional data into categorical and numerical variables; • processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables; • concatenation of processed additional variables, as well as the output value of the 1st MO model and construction of a vector based on them; • feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model; • feeding the output value of the 2nd MO model to the third MO model to form the output value of the second ensemble; - process the received signals using a third ensemble of machine learning models trained on a combination of topics, and during this processing the following is carried out: signal processing by the first MO model, during which the following is performed: • defining time intervals for retrieving variables based on the start and end timestamps of the question and the answer timestamp and based on the question type and topic; • extracting variables from each signal at certain time intervals; • processing of obtained variables from signals, which involves normalization and concatenation of processed variables and construction of a vector based on them; • feeding the mentioned vector into the 1st MO model to obtain the output value of the 1st MO model; • transfer of the output value of the 1st MO model to the input of the 2nd MO model; using the second MO model, the output value of the 1st MO model and additional data are processed, and during this processing the following is carried out: • dividing additional data into categorical and numerical variables; • processing of obtained variables from additional data, which involves vectorization of categorical variables and normalization of numerical variables; • concatenation of processed additional variables, as well as the output value of the 1st MO model and construction of a vector based on them; • feeding the mentioned vector into the 2nd MO model to obtain the output value of the 2nd MO model; • feeding the output value of the 2nd MO model to the third MO model to form the output value of the third ensemble; - using the third MO model, the output values of the first, second and third MO ensembles are processed, and during this processing the following is carried out: concatenation of the processed output values of the first, second and third ensembles and construction of a vector based on them; feeding said vector to the 3rd MO model to obtain the output value of the 3rd MO model; comparison of the output value of the 3rd model with a given threshold value; And - determine that the response is false if the output value is greater than or equal to the threshold value, or the response is true if the output value is below the threshold value. 2. The method according to claim 1, characterized in that the ML models are trained on one of the topics for checks or a combination thereof, where the topics for checks are: narcotic substances, receiving additional remuneration, disclosure of confidential information, debt obligations, third-party income, criminal offenses , administrative offenses, violations of internal regulatory documents (INR). 3. The method according to claim 1, characterized in that the MR models of the first, second and third ensembles have a type selected from the group: gradient boosting, random forest or neural network. 4. The method according to claim 1, characterized in that the third ML model has a type selected from the group: logistic regression, random forest, gradient boosting or averaging. 5. The method according to claim 1, characterized in that the printing test records contain signals from sensors including at least one of: heart rate (HR), galvanic skin response (GSR), blood pressure, upper and lower respiration, piezoplethysmogram, photoplethysmogram, thermal, pupil movements or combinations thereof. 6. The method according to claim 1, characterized in that the additional data contains at least one of: identification number of polygraph examiners, identification number of polygraphs, information about weather conditions, results of an electroencephalogram, magnetic resonance imaging, functional near-infrared spectroscopy, information about geomagnetic storms or combinations thereof. 7. A system for automatically checking a subject using a polygraph using machine learning methods, containing: - at least one processor; - at least one memory connected to the processor, which contains machine-readable instructions that, when executed by at least one processor, enable execution of the method according to any one of claims. 1-6.

Images