Method and system for recognizing information representing a trade secret

Abstract

A method for recognizing within a text document data that represents a trade secret includes segmenting text data into sentences, tokenizing the text data, adding a special token before each sentence, vectorizing said tokens, processing vector representations of the tokens using a neural network-based machine learning model trained on datasets containing a trade secret, and labelling a document containing a trade secret with a label indicating the presence of a trade secret. The technical result is that of providing more accurate recognition of data containing a trade secret.

Description

[0001] METHOD AND SYSTEM FOR RECOGNITION OF INFORMATION CONSTITUTING A COMMERCIAL SECRET

[0002] AREA OF TECHNOLOGY

[0003] [1] This technical solution generally relates to the field of data protection, and more specifically to the recognition of data constituting a commercial secret in text documents.

[0004] LEVEL OF TECHNOLOGY

[0005] [2] Currently, any organization has restrictions on the storage, processing and distribution of confidential data. Preventing the distribution of confidential information is an important issue in the field of information security, since its leakage can entail serious consequences. In this regard, the recognition of confidential information is one of the most important problems in this area of ​​technology.

[0006] [3] In addition, certain types of confidential information are additionally subject to the requirements for protection and handling established by law, regulating the processing of such information, for example, Federal Law No. 98-FZ of 29.07.2004 "On Commercial Secrets" (hereinafter referred to as FZ-98) regulates relations related to the establishment, change and termination of the commercial secret regime in relation to information that has actual or potential commercial value due to its being unknown to third parties. Thus, in accordance with the law, when implementing the commercial secret (CS) regime, it is necessary to organize a number of measures to protect confidential information, which include: restricting access to information constituting a commercial secret (ICT); accounting for persons who have gained access to ICT; regulation of relations on the use of ICT; affixing the "Commercial Secret" stamp on ICT media and including it in the composition of the details of documents containing ICT, including electronic ones.

[0007] [4] Due to the fact that the “Commercial Secret” classification is not always affixed by the data owner or is affixed incorrectly, it is necessary to identify documents that actually contain ICT and at the same time do not contain the “Commercial Secret” classification required in such cases, in order to fulfill the duties of the legal owner of ICT and to comply with the requirements of Article 10 of Federal Law 98: adding the “Commercial Secret” classification, restricting access to ICT and keeping records of persons to whom such access has been granted. At the same time, the main difficulty in recognizing information containing a commercial secret is the lack of specific categories of information related to ICT, which, accordingly, does not allow the use of existing approaches to searching for confidential data, due to their unsuitability for searching for information containing ICT, but not related to the standard categories of confidential data.

[0008] [5] Thus, an approach based on manual recognition of such information in documents is known from the prior art.

[0009] [6] The disadvantages of this approach include low accuracy of ICT recognition, due to the need for employees who understand the regulatory documents on commercial secrets. In addition, this approach is highly susceptible to the influence of the human factor.

[0010] [7] The prior art also includes a solution disclosed in Chinese Patent No. CN 111259116 A (Beijing Luoan Technology Co Ltd), published on 16.01.2020. This solution, in particular, discloses a method for detecting data constituting a trade secret based on the detection of sensitive data.

[0011] [8] The disadvantages of this solution are the low accuracy of detecting data that constitute ICT, due to the presence of ICT in data that is not confidential.

[0012] [9] Common shortcomings of known solutions include low accuracy and efficiency of recognizing information containing commercial secrets in text documents, due to the need to take into account the semantic context of the recognized data. In addition, such a solution should ensure the possibility of secure data transfer by automatically identifying documents containing ICT, but not having the appropriate classification.

[0013] DISCLOSURE OF INVENTION

[0014]

[0010] This technical solution is aimed at eliminating the shortcomings inherent in existing solutions known from the state of the art.

[0011] The technical problem solved in this technical solution is the creation of a new and effective method for recognizing data constituting a commercial secret in text documents.

[0015]

[0012] The technical result that is manifested in solving the above-mentioned problem is an increase in the accuracy of recognizing data containing a commercial secret, due to the use of the semantic context of the recognized data.

[0016]

[0013] The specified technical results are achieved by recognizing data constituting a trade secret in text documents, comprising the following stages: a) receiving a document containing text data; b) segmenting the text data obtained in stage a), wherein during the segmentation the text data is divided into sentences; c) tokenizing the text data, wherein during the tokenization a service token is added before each sentence; d) vectorizing the tokens obtained in stage c); e) processing the vector representations of the tokens obtained in stage d) using a machine learning model based on a neural network trained on data sets containing a trade secret, during which the following is performed: i. determining the proximity of each vector to the vectors of all other tokens in the sentence; ii. forming a vector for each token that takes into account the contextual relationship; iii.combining the generated vectors into a common sentence vector formed as a result of transforming the service token; iv. determining whether the common vector belongs to the category of a commercial secret; f) assigning a commercial secret label to the document containing the commercial secret.

[0014] In one particular example of implementing the method, the common vector is formed at least by means of an average or maximum concatenation algorithm.

[0017]

[0015] In another particular example of implementing the method, the category of commercial secret is at least one of the following categories: a) promising and strategic commercial plans, expansion plans; b) research and development programs, key ideas of research and development work.

[0018]

[0016] In another particular example of implementation, the method additionally comprises the step of determining whether a document is a commercial secret.

[0019]

[0017] In another particular example of implementing the method, the document’s belonging to a commercial secret is determined on the basis of the proportion of sentences containing a commercial secret relative to the total number of sentences in the document.

[0020]

[0018] In addition, the claimed technical results are achieved through a system for recognizing data constituting a commercial secret in text documents, containing:

[0021] • at least one processor;

[0022] • at least one memory connected to the processor, which contains machine-readable instructions that, when executed by at least one processor, ensure the execution of a method for recognizing data constituting a commercial secret in text documents.

[0023] BRIEF DESCRIPTION OF DRAWINGS

[0024]

[0019] The features and advantages of the present technical solution will become apparent from the following detailed description and the accompanying drawings, in which:

[0025] 120] Fig. 1 illustrates a block diagram of the implementation of the claimed method.

[0026]

[0021] Fig. 2 illustrates an example of the implementation of the attention mechanism.

[0027]

[0022] Fig. 3 illustrates an example of a computing device for implementing the claimed systems. IMPLEMENTATION OF THE INVENTION

[0028] [231 The claimed technical solution proposes a new approach, providing a new and effective method of recognizing data containing a commercial secret in text documents. The main feature of the claimed solution is the ability to recognize data containing a commercial secret in text documents, due to the implementation of an internal attention mechanism, which allows taking into account the contextual relationships between words and sentences of a text file. In addition, the implementation of this technical solution increases the accuracy of recognizing the presence of a commercial secret in text files, due to the implementation of a machine learning model based on a neural network, providing a search for complex dependencies and the ability to recognize phrases containing a commercial secret that have not previously been encountered in documents.Also, another additional advantage achieved by using the declared solution is the ability to ensure secure data transfer, due to the automatic detection of documents containing ICT, but not having the appropriate classification, and, accordingly, the application of data transfer restriction policies to such documents in real time.

[0029]

[0024] The claimed technical solution may be implemented, for example, by a system, a machine-readable medium, a server, etc. In this technical solution, a system means, among other things, a computer system, a computer (electronic computer), a CNC (numerical control), a PLC (programmable logic controller), computerized control systems and any other devices capable of performing a given, clearly defined sequence of operations (actions, instructions).

[0030]

[0025] A command processing unit is an electronic unit or integrated circuit (microprocessor) that executes machine instructions (programs).

[0031]

[0026] The command processing unit reads and executes machine instructions (programs) from one or more data storage devices, such as random access memory (RAM) and / or read-only memory (ROM). ROM may include, but is not limited to, hard disk drives (HDD), flash memory, solid-state drives (SSD), optical storage media (CD, DVD, BD, MD, etc.), etc.

[0027] A program is a sequence of instructions intended for execution by a control unit of a computing machine or a command processing unit.

[0032]

[0028] The term "instructions" as used in this application may generally refer to software instructions or software commands that are written in a given programming language to perform a specific function, such as, for example, receiving and processing data, forming a user profile, receiving and transmitting signals, analyzing received data, identifying a user, etc. The instructions may be implemented in a variety of ways, including, for example, object-oriented methods. For example, the instructions may be implemented using the C++ programming language, Java, Python, various libraries (e.g., "MFC"; Microsoft Foundation Classes), etc. The instructions that perform the processes described in this solution may be transmitted via both wired and wireless data transmission channels, such as Wi-Fi, Bluetooth, USB, WLAN, LAN, etc.

[0033]

[0029] Fig. 1 shows a block diagram of a method 100 for recognizing data constituting a commercial secret in text documents, which is disclosed step by step in more detail below. The said method 100 consists of performing steps aimed at processing various digital data. The processing is typically performed using a system that can represent, for example, a server, a computer, a mobile device, a computing device, etc., supplemented by a hardware and / or software and hardware unit implementing the present method 100. The elements of the system are disclosed in more detail in Fig. 3.

[0034]

[0030] In this decision, the term “trade secret” or “information constituting a trade secret” (ICT) shall be understood to mean data of any nature (production, technical, economic, organizational, and others), including data on the results of intellectual activity in the scientific and technical sphere, as well as data on the methods of carrying out professional activities that have actual or potential commercial value due to their being unknown to third parties, to which third parties do not have free access on a legal basis, and in relation to which the owner of such information has introduced a trade secret regime.Thus, data constituting a commercial secret may not contain data access to which is restricted in accordance with the security policies of organizations and / or legislative acts, such as personal data, information constituting a banking secret, for example, full names of employees and clients, partners and suppliers, addresses, telephone numbers, email addresses, social security numbers, bank card information, TIN number, vehicle registration number, BIC number, IP address, geolocation data, marriage document number, education document number, date, URL address, MAC address, work book number, military ID number, OKPO code, etc., without limitation, but at the same time be of critical importance to the organization, for example, data on the company's development plans, data on the company's structure, etc.

[0035]

[0031] As indicated above, the difficulty of detecting data containing commercial secrets in text documents is the lack of clear categories of data that relate to CT, due to which the systems known from the state of the art, aimed at recognizing confidential data, are of little use and / or are not capable of recognizing such data at all. For example, data containing CT may include prospective and strategic plans of companies, and for each specific company and / or organization, such data may have completely different categories, i.e., they may differ significantly from each other.

[0036]

[0032] The present technical solution is intended to solve the above-mentioned problems.

[0037]

[0033] At step 110, a document containing text data is received.

[0038]

[0034] At the specified step 110, a document containing text data is received by the system, for example, the system 200. In one particular embodiment, the document can be loaded into the system 200 via a communication network, such as the Internet, LAN, etc. In another particular embodiment, the document can be imported directly from a flash drive and / or built-in memory of the system 200. In this solution, a document containing text data should be understood as any data file containing text data. Thus, the document can be any unstructured document, for example, a Word file, PDF, text document, photograph, digital document, e-mail file, etc.

[0039]

[0035] The method 100 then proceeds to step 120.

[0036] At step 120, the text data obtained at step 110 is segmented, wherein during the segmentation, the text data is broken down into sentences.

[0040]

[0037] The specified step 120 can be performed, for example, by searching for technical symbols in the text (for example, a period, an exclamation mark, a question mark, etc.), taking into account the division of said symbols into groups, etc. In another particular embodiment, the text can be broken down into phrases, which can be expressed by sentences, phrases, paragraphs or individual words. For example, the text "JSC Bank plans to acquire a key stake in the organization JSC Firm in order to take a larger share of the market. The details of the transaction are being discussed." will be broken down into the sentences "JSC Bank plans to acquire a key stake in the organization JSC Firm in order to take a larger share of the market." and "The details of the transaction are being discussed." The choice between the level of text segmentation is determined by how much detail is needed to understand where the CT is located in the text. The most acceptable level is segmentation into sentences.

[0041]

[0038] Next, phrases are cleared of service characters and final punctuation marks.

[0042]

[0039] For example, the sentence “JSC Bank plans to acquire a key stake in the organization JSC Firm in order to occupy a larger share of the market.” will be cleared of the end point and will take the form “JSC Bank plans to acquire a key stake in the organization JSC Firm in order to occupy a larger share of the market.”

[0043]

[0040] Extracting text from a document in this solution should be understood as extracting certain parts of the text from a text document and storing said parts in the memory of the system 200, for example, in the form of a data file. That is, at the indicated stage 120, the original text is converted into several parts, in accordance with the segmentation level. In another particular embodiment, as a tool for determining the beginning and end of a part (phrase, sentence), both internal document tools that provide a character count in the text, for example, text editors, and tools, for example, software tools for analyzing text documents, including those built into the system 200, can be used.

[0044]

[0041] Thus, at the specified step 120, segmentation of text data is performed.

[0045]

[0042] The method 100 then proceeds to step 130.

[0043] At step 130, tokenization of the text data is performed, wherein during the tokenization, a service token is added before each sentence.

[0046]

[0044] At step 130, the received text data is processed. The input text is tokenized, i.e. segmented into parts, for example, into sentences, words or symbols. In this solution, a token is understood to be a sequence of symbols in a document that is important for analysis. Thus, tokens can be, for example, individual words, parts of words, etc. Tokenization into sentences can be carried out using, for example, lexical analyzers such as razdel, rusenttokenize, NLTK, etc. Lexical analysis is the process of analytically parsing the input sequence of symbols into recognized groups (lexemes) in order to obtain identified sequences called "tokens" at the output. In addition, tokenization of the input text can be carried out based on regular expressions.It is obvious to a person skilled in the art that any lexical analyzer known from the state of the art can be used and this solution is not limited to the examples given above. It is worth noting that in one particular embodiment, the token is assigned a position in the text and a corresponding ordinal number.

[0047]

[0045] The main feature of this step 130 is that a service token is added before each sentence received in step 120. Thus, in one particular embodiment, a service token is added before each phrase received in step 120.

[0048]

[0046] The service token is a fixed numerical value obtained from the trained token table containing the numerical representation for verbal tokens, since the model works with numbers. At the same time, the service token as such does not have a semantic interpretation in any language. The specified token is necessary for obtaining the embedding of the entire phrase and its classification. Also, the model configurations indicate which token will be the service token, i.e. the summary token for the internal attention mechanism. Thus, each token is replaced with its identifier from the token table of the trained model. In addition, the specified service token is necessary for the correct operation of the attention mechanism, which takes into account the contextual relationships between words within a phrase / sentence.

[0049]

[0047] At step 140, the tokens obtained at step 130 are vectorized.

[0048] At said step 140, each token obtained in the tokenization process is vectorized, for example, using embeddings or direct encoding (one hot encoding). Thus, for example, during tokenization, each token is represented in the dictionary by its index, which reflects the position in the said dictionary. Thus, each token represents an index in the dictionary, and, accordingly, the vectorization process is carried out by replacing each token with its index in the dictionary. Then, the indices are grouped taking into account the sparseness of the dictionary and the semantic proximity of the tokens. It will be obvious to a person skilled in the art that other vectorization algorithms can also be used for vectorizing tokens, for example, using the TransformersBertEmbedder, Word2vec, fastText algorithms, etc., without limitation.The specified vectorization process is a preparatory step for processing the data by a machine learning model that performs recognition of sensitive data (step 150).

[0050]

[0049] Thus, in another particular embodiment, the embeddings of BERT, GLoVe, ELMo, word2vec and other neural network architectures, TF-IDF vectors, etc., can be presented as vectors, without being limited to the examples given. It is worth noting that the token embedding takes into account its value and position in the text.

[0051]

[0050] Method 100 then proceeds to step 150

[0052]

[0051] At step 150, vector representations of tokens are processed using a machine learning model based on a neural network trained on data sets containing commercial secrets, during which the belonging of each token to the category of confidential data is determined.

[0053]

[0052] A machine learning model was developed and applied to process unstructured data and identify data containing commercial secrets in unstructured text. The peculiarity of this machine learning model is the attention mechanism. Thus, with the help of a neural network and an attention mechanism, it is possible to analyze the embeddings of all tokens at once in all directions, due to which the machine learning model extracts from a phrase not just categorized data, but contextual relationships between words. Processing text in all directions allows taking into account the context of words relative to both previous and subsequent words. The attention mechanism, in turn, provides the ability to highlight more significant tokens in the entire phrase and take into account the relationship of each token with each other token in the entire phrase.In this way, new token embeddings are formed that take into account the context, including the embedding of a service token that takes into account the context of the entire sentence at once.

[0054]

[0053] The ML model was trained on pre-labeled data. At the time of model creation, a dataset of labeled confidential data consisting of over 100 thousand phrases represented by 2.5 million tokens was used. The dataset was formed based on original text documents containing personal data, information constituting a banking and commercial secret, etc. Training the model to recognize confidential data in text documents consisted of identifying sentence classes in the text, i.e. commercial secret categories (a) promising and strategic commercial plans; b) research and development programs, etc.). When labeling the data, each sentence represented by a set of tokens is assigned a tag corresponding to the commercial secret category. If the sentence is not a commercial secret, a zero tag is assigned.The dataset can be represented, for example, as an ordered list of tokens grouped by sentences. It is obvious to a specialist in this field of technology that any dataset markup and presentation scheme known from the state of the art can be applied, and this solution is not limited to the examples given above. 80% of the tokens from the dataset were used to train the ML model, and 20% were used to calculate the quality metrics. Training was performed on 5 epochs.

[0055] 154] The main metric of the ML model quality, the F1 measure weighted by classes, is about 93%. At the same time, the spread between recall and precision does not exceed 0.8%, which indicates a high level of model quality both from the point of view of security - low probability of missing confidential information, and from the point of view of the end user - a small number of false detections of confidential information. The degree of drop in the quality metric during text augmentation at the character level, showing the stability of the model, does not exceed 5%.

[0056]

[0055] The said neural network-based machine learning model was successfully implemented and tested in the organization in the work processes of the departments.

[0056] Let us consider in more detail the principle of operation of the attention mechanism in the neural network-based machine learning model. Thus, Fig. 2 shows an example of the operation of the model with the attention mechanism. In one particular embodiment, for example, a neural network of the Transformer architecture with a multi-head self-attention mechanism can be used as a neural network.

[0057]

[0057] In more detail, the mechanism of attention is described by the formula attention =

[0058] ( QK softmax I - = 1

[0059] V, where Q is the query vector, K is the key vector, V is the value vector, dk is

[0060] \Vdfc. vector dimension.

[0061]

[0058] The vectors Q, K, and V are obtained by multiplying the token embedding by the corresponding matrices obtained during the preliminary training of the model. Since the calculations are actually performed on the vector representations of several tokens, Q, K, and V are matrices, and the matrix K must be transposed before calculating the product of Q and K. The key vector and the value vector serve to represent the token, and the query vector shows the importance of a given token relative to others. z i

[0062] The softmax function is expressed by the formula tr(Z)i = : — tt, where 1 and j are the indices of the element =i. е J vector in the range from 1 to N, serves for normalization, i.e. transforms a vector Z of dimension N to a vector of the same dimension, where all coordinates of the normalized vector щ are expressed by a number in the range from 0 to 1, and their sum is equal to one.

[0063]

[0059] Token embeddings that take into account contextual relationships are combined into a common embedding of the entire phrase, represented by the embedding of the service token. This embedding can be formed using various algorithms, for example, by concatenating mean pooling and max pooling. Max pooling is obtained by choosing the embedding of the token with the maximum value in the entire phrase, and mean pooling is obtained by averaging the embeddings of all tokens in the phrase.

[0064]

[0060] Next, using a neural network trained on sets of texts containing a commercial secret and phrase embedding (embedding of a service token), the model assigns the phrase to a specific item in the ICT list. For example, 1.1. Prospective and strategic commercial plans, expansion plans or 1.2. Research and development programs, key ideas of research and development work. If the phrase does not relate to ICT, then a zero tag is entered.

[0065]

[0061] For example, continuing the example discussed above, the resulting phrase would be classified as follows:

[0066] Table 1

[0067]

[0062] Thus, to take into account the contextual relationships between words when searching for CT, after breaking the text into words (tokens), their vectors (embeddings) are calculated. Then the vector of each word is compared with the vectors of all other words in the sentence and their proximity is analyzed, thereby extracting the contextual relationships between words. Due to this, each word gets a new vector that takes into account the context. Then the vectors of all words are combined into one common vector, essentially containing the context of the entire sentence.

[0068]

[0063] In one particular embodiment, the said sequence of actions is repeated a specified number of times. This approach allows for taking into account various interrelations between words. Thus, in another particular embodiment, the said approach is repeated 12 times with different pre-trained weights. It is obvious to a person skilled in the art that the number of repetitions can take any positive integer and depends on the number of layers of the neural network. Continuing to consider this feature, with each repetition, the new sentence vector takes into account the interrelations of the current layer and all previous ones. Thus, at the output we have a sentence vector that takes into account the context for various interrelations of words.

[0069]

[0064] The proposed approach to taking into account contextual relationships of words allows us to determine with high accuracy the relationships between adjacent words in a sentence.

[0070]

[0065] In addition, in another particular embodiment, it is possible to isolate specific sentences containing a commercial secret, for example, for subsequent depersonalization. That is, the ability to determine precisely sentences, and not simply classify the entire document as a whole, provides deeper control over the transmitted data and makes it possible to depersonalize data containing a commercial secret without preventing the transmission of the entire document.

[0071]

[0066] In addition, in another particular embodiment, after determining the data containing CT, said data can be extracted for further processing, for example, into the secure memory of the system 200 for further anonymization.

[0072]

[0067] Returning to the steps of step 150, the first step involves determining the proximity of each vector to the vectors of all other tokens in the sentence.

[0073]

[0068] As noted above, to determine the proximity of each vector to the vectors of other tokens in the sentence, vectors Q, K, and V are determined. The key vector K and the value vector V serve to represent the token, and the query vector Q shows the significance of this token relative to others. The vectors are taken from the model matrices formed during its training. The softmax function serves for normalization, i.e. it transforms a vector Z of dimension N to a vector o of the same dimension, where all coordinates of the normalized vector o are expressed as a number in the range from 0 to 1, and their sum is equal to one. The closer to 1 the calculated by

[0074] Q K I \ softmax formula V the weight of the token vector, the greater the importance it has on the token vector relative to which the calculations are performed, and the closer these tokens are in terms of context. The new token vector, taking into account the context, is formed by summing the products of the weights and values ​​of the remaining token vectors of the phrase.

[0075]

[0069] For example, in the phrase "Strategic plan to buy a promising company and sell an unprofitable organization", the vector of the token "purchase" has a greater weight relative to the vector of the token "company" than relative to the vector of the token "organization". Thus, the attention mechanism determines that we are talking about buying a company, not an organization. By summing up all the relationships between the vectors of the phrase's tokens, a vector representation of the context is formed, on the basis of which the phrase's belonging to a commercial secret is determined.

[0076]

[0070] Next, in step two, the vectors obtained in step one are cyclically passed through the layers of the neural network, where they are subjected to transformations similar to step one. The presence of several layers of the neural network with an attention mechanism allows for the extraction of more features that influence the contextual relationships of the tokens. Thus, for example, in one particular embodiment, in said step two, twelve repetitions may be performed, for example.

[0077]

[0071] In step three, the generated vectors are combined into a common sentence vector. This common vector is an embedding of the service token that takes into account the context of the entire sentence. The token embeddings that take into account contextual relations are combined into a common embedding of the entire phrase, represented by the embedding of the service token. This embedding can be formed using various algorithms, for example, by concatenating the average (mean pooling) and maximum pooling (max pooling). Max pooling is obtained by selecting the embedding of the token with the maximum value in the entire phrase, and mean pooling is obtained by averaging the embeddings of all tokens in the phrase. Continuing the example from step 1, a common embedding of the entire phrase (a vector representation of the service token) is formed that takes into account the context that this phrase is about promising plans and expansion.

[0078]

[0072] In step four, the general vector is determined to belong to the category of commercial secret.

[0079]

[0073] Next, using a neural network trained on sets of texts containing a commercial secret and phrase embedding (embedding of a service token), the model assigns the phrase to a specific item of the ICT list. For example, 1.1. Promising and strategic commercial plans, expansion plans or 1.2. Research and development programs, key ideas of research and development work. If the phrase does not relate to ICT, then a zero tag is entered. The neural network classifier that determines whether a phrase belongs to a specific item of the ICT list can be represented by a linear classifier, a conditional random field (CRF), or any other solution known from the technical field. Continuing the example, a classifier based on embedding containing the context of promising expansion plans assigns the phrase to item 1.1. Prospective and strategic commercial plans, plans to expand the range of ICT.

[0080]

[0074] Thus, at the specified step 150, phrases / sentences containing a commercial secret are recognized and extracted from the text document. In another particular embodiment, phrases or sentences containing a commercial secret are assigned a category of a commercial secret, as was specified above.

[0075] Next, at step 160, a document containing a commercial secret is assigned a label of the presence of a commercial secret.

[0081]

[0076] Thus, at the indicated step 160, based on the original document and phrases with the corresponding tags, set as a result of processing the document at step 150, a report is generated in which phrases containing ICT are highlighted. The indicated report can be generated, for example, in a graphical user interface. In another particular embodiment, a document containing a commercial secret is assigned a label (stamp) - commercial secret.

[0082]

[0077] Thus, the specified label can characterize the permitted channels of data transmission within the organization, the circle of persons having access to such data and other actions with documents, conditioned by the protection of such a document from falling into the hands of third parties.

[0083]

[0078] In one particular embodiment, after a document containing ICT is identified, a digital imprint (hash) is taken from the document to assign a label to said document. Said imprint contains information about the document, such as the presence of ICT, the owner of the document (user), etc. Accordingly, said hash is then collected, saved and processed centrally, in the company's security system. Thus, when transmitting any document within the network, including a document containing ICT, its hash is requested by the security system and compared with the list of hashes of documents containing ICT. Said system, in particular, ensures access control, transmission, encryption, i.e. ensures secure transmission and processing of such a document. In addition, thanks to such a security system, a policy of ICT regime management and compliance with the requirements of Article 10 of Federal Law-98 (person registration, classification, access restriction) is implemented.

[0084]

[0079] Thus, in another particular embodiment, the said solution can be built into the organization's data transmission channels. Accordingly, before sending documents, the said documents will be checked for the presence of data containing a commercial secret. If the fact of an attempt to transmit such data containing an ICT label is recorded, then the transmission of the document is blocked.

[0085]

[0080] Thus, the above materials described a method for recognizing data containing commercial secrets in text documents.

[0081] Now let us consider a scenario for the application of some variants of the declared solution.

[0086]

[0082] Thus, one of the application scenarios may be the identification of data containing a commercial secret. As indicated above, the complexity of such a procedure lies in the fact that a commercial secret may be data that is not sensitive as such. For example, the sentence "The bank plans to take over a key metallurgical company." As can be seen, this sentence does not contain data protected by the law on personal data or on banking secrecy, but at the same time, based on the context, this expression refers to a commercial secret, since it contains promising and strategic commercial plans, plans for expansion. That is, a feature of the specified method 100 is the ability to recognize the specified data, even in the absence of sensitive data in the document.

[0087]

[0083] Continuing with the example, a system such as system 200 may be integrated into an organization's transmission channels, such as a mail client, etc. Accordingly, all documents transmitted through such a channel pass through the system. In order to determine the presence of data containing a commercial secret, as well as to prevent the dissemination of such data, the document is segmented into sentences (step 120), tokenized, and during the tokenization, a service token is added before each sentence (step 130). Next, the tokens are vectorized (step 140) and processed (step 150).

[0088]

[0084] As indicated above, one of the features of the claimed technical solution is the use of an attention mechanism that provides the ability to take into account contextual relationships of sentence tokens (semantic connection), which allows for increased accuracy of ICT recognition in documents, including through the ability to identify sentences, as shown in the example above. To do this, at the first step of data processing, the proximity of each vector to the vectors of all other tokens in the sentence is determined. Thus, at this step, the proximity of each vector to the others within the sentence is calculated. It is worth noting that this action is performed together with the service token. For example, if a sentence consists of two words, it will have the following structure [cis] tl t2, where [cis] is the vector of the service token, and tl and t2 are the vector representation of the sentence tokens. Accordingly, for each vector representation of the token, its proximity to the others is calculated.In this case, the vectors of the token tl in relation to [cis] and t2, the vectors of the token t2 in relation to [cis] and tl, and the vectors of the token [cis] in relation to tl and t2. The result of such a calculation will contain a new vector for each token, constructed taking into account the proximity of the specified token to the others. Accordingly, in step two, the specified procedure is cyclically repeated a specified number of times, for example, twelve, for the resulting new vectors, and in each step, the vectors from the previous step are taken. It is worth noting that cyclic repetition is necessary to ensure that all relationships (semantic connections / contextual relationships) between words are taken into account. Next, in step three, the resulting vectors are combined into a common sentence vector, on the basis of which, in step four, the belonging of the common vector to the category of a commercial secret is determined.

[0089]

[0085] Returning to the given example, if during the processing of a document, the data contains ICT, for example, one of the above-mentioned ICT categories, then the document is assigned a digital label. Said label is transmitted to the security system. In addition, in another particular embodiment, if the document contains ICT, but was not initially marked with the appropriate classification, but was detected during the analysis process, then the system, such as system 200, blocks the transmission of such a document. In another particular embodiment, a notification is sent to the user about the inadmissibility of transmitting such information. In addition, as indicated above, on the basis of this label, further distribution and processing of the document is ensured.

[0090]

[0086] In another particular embodiment, the present method 100 can implement recognition of specific sentences in the text containing ICT, i.e., classify the text by sentences. When using such an approach, the proportion of sentences with ICT relative to the total number of sentences can also be calculated. Based on this proportion, a false triggering of the algorithm can be determined. Thus, for example, when identifying data containing ICT in 1 sentence, with a total number of sentences of 2000, a decision will be made about a false triggering of the algorithm and the document will be categorized as not containing ICT. On the contrary, if 9 sentences out of 10 (90%) with ICT are recognized in the document being processed, then the document will be considered as containing ICT. The indicated features also make it possible to increase the accuracy of ICT recognition in text documents.

[0087] Fig. 3 shows a system 200 implementing the steps of the claimed method 100.

[0091]

[0088] In general, the system 200 comprises components such as: one or more processors 201, at least one memory 202, data storage means 203, input / output interfaces 204, I / O means 205, network interaction means 206, which are connected via a universal bus.

[0092]

[0089] The processor 201 performs the basic computing operations necessary for processing the data when performing the method 100. The processor 201 executes the necessary machine-readable instructions contained in the RAM 202.

[0093]

[0090] Memory 202 is typically in the form of RAM and contains the necessary software logic to provide the required functionality.

[0094]

[0091] Data storage means 203 can be implemented in the form of HDD, SSD disks, RAID array, flash memory, optical storage devices (CD, DVD, MD, Blue-Ray disks), etc. Means 203 allow long-term storage of various types of information, for example, the history of the substitution table, the formatting table, etc.

[0095]

[0092] To organize the operation of the components of the system 200 and to organize the operation of external connected devices, various types of I / O interfaces 204 are used. The selection of the corresponding interfaces depends on the specific design of the computing device, which may be, without limitation: PCI, AGP, PS / 2, IrDa, FireWire, 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.

[0096]

[0093] The selection of interfaces 204 depends on the specific implementation of system 200, which can be implemented on the basis of a wide range of devices, for example, a personal computer, mainframe, laptop, server cluster, thin client, smartphone, server, etc.

[0097]

[0094] The following may be used as I / O data means 205: a keyboard, a joystick, a display (touch display), a monitor, a touch display, a touchpad, a mouse, a light pen, a stylus, a touch panel, a trackball, speakers, a microphone, augmented reality means, optical sensors, a tablet, light indicators, a projector, a camera, biometric identification means (retina scanner, fingerprint scanner, voice recognition module), etc.

[0095] Network interaction means 206 are selected from devices that provide network data reception and transmission, for example, an Ethernet card, a WLAN / Wi-Fi module, a Bluetooth module, a BLE module, an NFC module, an IrDa, an RFID module, a GSM modem, etc.By means of means 606, the organization of data exchange is ensured between, for example, system 200, presented in the form of a server, and the user's computing device, on which the received data (anonymized text document) can be displayed via a wired or wireless data transmission channel, for example, WAN, PAN, LAN, Intranet, Internet, WLAN, WMAN or GSM.

[0098]

[0096] The submitted application materials disclose preferred examples of the implementation of the 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.

Claims

FORMULA 1. A method for recognizing data constituting a trade secret in text documents, comprising the steps of: a) receiving a document containing text data; b) segmenting the text data obtained in step a), wherein the text data is divided into sentences during the segmentation; c) tokenizing the text data, wherein a service token is added before each sentence during the tokenization; d) vectorizing the tokens obtained in step c); e) processing the vector representations of the tokens obtained in step d) using a machine learning model based on a neural network trained on data sets containing a trade secret, during which the following are performed: i. determining the proximity of each vector to the vectors of all other tokens in the sentence; ii. forming a vector for each token that takes into account the contextual relationship; iii.combining the generated vectors into a common sentence vector formed as a result of transforming the service token; iv. determining whether the common vector belongs to the category of a commercial secret; f) assigning a commercial secret label to a document containing a commercial secret.

2. The method according to claim 1, characterized in that the common vector is formed by at least a mean concatenation or a max union algorithm.

3. The method according to paragraph 1, characterized in that the category of commercial secret is at least one of the following categories: i. promising and strategic commercial plans, expansion plans; ii. research and development programs, key ideas of research and development work.

4. The method according to paragraph 1, characterized in that it additionally contains the step of determining whether a document is a commercial secret.

5. The method according to paragraph 4, characterized in that the document’s classification as a commercial secret is determined on the basis of the proportion of sentences containing a commercial secret relative to the total number of sentences in the document.

6. A system for recognizing data constituting a commercial secret in text documents, 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, ensure the execution of the method according to any one of claims 1-5.

Images