RU2673708C1 - System and method of machine training model of detecting malicious files - Google Patents

Abstract

FIELD: information technology.SUBSTANCE: invention relates to anti-virus technologies, namely the detection of malicious files. Method of machine learning the malware detection model by using the machine learning model to detect a malicious file method of learning a detection model that provides monotonous changes in the severity of the file depending on the change in the number of patterns of behavior of the malicious file generated based on the analysis of the behavior log.EFFECT: providing the teaching on the detection model for malicious files.10 cl, 5 dwg

Description

Technical field

The invention relates to antivirus technologies, and more particularly to systems and methods for detecting malicious files.

State of the art

The rapid development of computer technology in the last decade, as well as the widespread dissemination of a variety of computing devices (personal computers, laptops, tablets, smartphones, etc.) have become a powerful incentive for using these devices in various fields of activity and for a huge number of tasks (from the Internet surfing to bank transfers and electronic document management). In parallel with the increase in the number of computing devices and software running on these devices, the number of malicious programs has grown rapidly.

Currently, there are a huge number of varieties of malware. Some of them steal personal and confidential data from users' devices (for example, logins and passwords, bank details, electronic documents). Others form so-called botnets from user devices for attacks such as denial of service (DDoS) or bruteforce bruteforce passwords to other computers or computer network. Still others offer paid content to users through intrusive ads, paid subscriptions, sending CMC to paid numbers, etc.

To combat malware, which includes the detection of malware, the prevention of infection and the restoration of the performance of computing devices infected with malware, specialized programs are used - antiviruses. To detect the whole variety of malicious programs, anti-virus programs use a variety of technologies, such as:

статический анализ - анализ программ на вредоносность, исключающий запуск или эмуляцию работы анализируемых программ, на основании данных содержащихся в файлах, составляющих анализируемые программы, при этом при статистическом анализе могут использоваться:

static analysis - analysis of programs for malware, excluding the launch or emulation of the work of the analyzed programs, based on the data contained in the files that make up the analyzed programs, while statistical analysis can be used:

сигнатурный анализ - поиск соответствий какого-либо участка кода анализируемых программ известному коду (сигнатуре) из базы данных сигнатур вредоносных программ;

Signature analysis - search for correspondence of any part of the code of the analyzed programs to the known code (signature) from the malware signature database;

белые и черные списки - поиск вычисленных контрольных сумм от анализируемых программ (или их частей) в базе данных контрольных сумм вредоносных программ (черные списки) или базе данных контрольных сумм безопасных программ (белые списки);

white and black lists - search for calculated checksums from the analyzed programs (or their parts) in the malware checksum database (black lists) or the safe checksums database (white lists);

динамический анализ - анализ программ на вредоносность на основании данных, полученных в ходе исполнения или эмуляции работы анализируемых программ, при этом при динамическом анализе могут использоваться:

dynamic analysis - analysis of malware programs on the basis of data obtained during the execution or emulation of the work of the analyzed programs, while dynamic analysis can use:

эвристический анализ - эмуляция работы анализируемых программ, создание журналов эмуляции (содержащих данные по вызовам API-функций, переданным параметрам, участкам кода анализируемых программ и т.д.) и поиск соответствий данных из созданных журналов с данными из базы данных поведенческих сигнатур вредоносных программ;

heuristic analysis - emulation of the work of the analyzed programs, the creation of emulation logs (containing data on calls to API functions, passed parameters, code sections of the analyzed programs, etc.) and the search for data matches from the created logs with data from the database of malware malware signature;

проактивная защита - перехват вызовов API-функций запущенных анализируемых программ, создания журналов поведения анализируемых программ (содержащих данные по вызовам API-функций, переданным параметрам, участкам кода анализируемых программ и т.д.) и поиск соответствий данных из созданных журналов с данными из базы данных вызовов вредоносных программ.

proactive defense - interception of calls to API functions of running analyzed programs, creation of behavior logs of analyzed programs (containing data on calls to API functions, passed parameters, code sections of analyzed programs, etc.) and search for data matches from created logs with data from the database malware call data.

Both static and dynamic analysis have their pros and cons. Static analysis is less demanding on the resources of the computing device on which the analysis is performed, and since it does not require execution or emulation of the analyzed program, statistical analysis is faster, but less efficient, i.e. it has a lower percentage of malware detection and a higher percentage of false positives (i.e. deciding on the harmfulness of a file analyzed by the anti-virus program, even though the analyzed file is safe). Dynamic analysis due to the fact that it uses the data obtained during the execution or emulation of the work of the analyzed program, is slower and makes higher demands on the resources of the computing device on which the analysis is performed, but at the same time it is more efficient. Modern antivirus programs use complex analysis, which includes both elements of static and dynamic analysis.

Since modern standards of information security require a rapid response to malicious programs (especially new ones), automatic malware detection tools come to the fore. For the effective operation of the mentioned tools, artificial intelligence elements and various methods of machine learning malware detection models are often used (i.e., a set of rules for deciding whether a file is harmful based on a certain set of input data describing a malicious file), which can effectively detect not only well-known malware or malware with well-known malicious behavior, but also new malware with unknown or poorly researched malware behavior, and quickly adapt (to learn) to detect new malware.

US Pat. No. 9,282,220 B2 describes a technology for detecting malware in network traffic. For this purpose, characteristic features are selected from data selected from network traffic (features characterizing the type of the executable file, the behavior of the executable file, the type of data transmitted over the computer network, for example, the type and size of data transmitted over the computer network, commands executed during file execution, the presence of predefined signatures in the file, etc.), which can be a characteristic description of the selected data, i.e. vector (eng. feature vector), composed of values corresponding to a certain set of features for the object containing the selected data. Using the models for detecting safe files, detecting malicious files, and determining the types of malicious files that have been pre-trained using machine learning methods based on patterns made up of characteristics similar to those mentioned above, determine the weight and type of malware selected data belongs to, and make a decision to detect malware in network traffic.

Although the technology described above copes well with the detection of malicious files that have some characteristic features (i.e., data describing some features of files from a certain set of files, for example, the presence of a graphical interface, data encryption, data transfer over a computer network, etc. ), similar to the characteristic features of already known malicious files, it is not able to cope with the detection of malicious files that have distinct characteristic features (albeit similar behavior) from the characteristic signs of already known malicious files, in addition, the technology described above does not disclose such aspects of machine learning models as testing and retraining models, as well as the formation and reformation (depending on the results of the above testing) of characteristic features.

The present invention allows to solve the problem of machine learning a model for detecting malicious files.

Disclosure of invention

The invention is intended for antivirus scanning of files.

The technical result of the present invention is to provide training model detection of malicious files.

Another technical result of the present invention is to reduce errors of the first and second kind when detecting malicious files using a trained model for detecting malicious files by using a machine learning model for detecting a malicious file, a method for teaching the detection model that provides a monotonic change in the degree of maliciousness of the file depending on the change in the number behavior patterns generated based on the analysis of the behavior log.

These results are achieved using a machine learning system for detecting malicious files, which contains a tool for generating behavior patterns designed to: generate at least one behavior pattern based on executable commands and parameters selected from the behavior log, while the behavior log is a collection executable commands (hereinafter referred to as the command) from the file, where each command corresponds to at least one parameter describing the command (further - setting), the behavior pattern is a set of at least one command and such a parameter, which describes all the commands of said set; transmitting the generated patterns of behavior to the means of forming convolution functions; means for generating convolution functions, intended for: forming such a convolution function from a behavior pattern such that the inverse convolution function from the result of such a convolution function over the obtained behavior pattern will have a degree of similarity with the obtained behavior pattern greater than a given value; transmitting the generated convolution function to the machine learning detection model; detection model machine learning tool for: learning a detection model, in which the parameters of the detection model are calculated using the obtained convolution function over the obtained behavior patterns, where the detection model is a set of rules for calculating the degree of maliciousness of a file based on at least one behavior pattern using calculated parameters of said detection model.

In another particular case of system implementation, behavior logs are pre-generated from at least two files, one of which is a safe file, and the other is a malicious file.

In another particular case of system implementation, machine learning of a previously created untrained detection model is performed.

In another particular case of implementing the system, at least: gradient boosting on decision trees serves as a method of machine learning for the detection model; decisive trees; nearest neighbors method kNN; support vector method.

In another particular case of the system implementation, a method for training the detection model is selected that ensures the monotonicity of the change in the degree of harmfulness of the file depending on the change in the number of behavior patterns generated based on the analysis of the behavior log.

These results are achieved by using the machine learning method of the malware detection model, the method comprising the steps that are implemented using the machine learning system of the malware detection model and in which: at least one behavior pattern is generated based on the executable commands and parameters selected from the behavior log, while the behavior log is a set of executable commands (hereinafter - the command) from the file, where each command corresponds to enshey least one parameter describing said command (hereinafter - the parameter), the behavior pattern is a set of at least one command and such a parameter, which describes all the commands of said set; form such a convolution function from the behavior pattern that the inverse convolution function from the result of such a convolution function over the generated behavior pattern will have a degree of similarity with the generated behavior pattern greater than the specified value; teach a detection model in which the parameters of the detection model are calculated using the generated convolution function over the obtained behavior patterns, where the detection model is a set of rules for calculating the degree of maliciousness of a file based on at least one behavior pattern using the calculated parameters of the mentioned detection model.

In another particular case of the method, the behavior logs are pre-generated from at least two files, one of which is a safe file, and the other is a malicious file.

In another particular case of the method, machine learning of a previously created untrained detection model is performed.

In another particular case of the implementation of the method, at least: gradient boosting on decision trees serves as a machine learning method for the detection model; decisive trees; nearest neighbors method kNN; support vector method.

In another particular case of the method implementation, a method for training a detection model that monotonously changes the degree of harmfulness of a file depending on changes in the number of behavior patterns generated based on an analysis of the behavior log is selected.

Brief Description of the Drawings

FIG. 1 is a block diagram of a machine learning system for detecting malicious files.

FIG. 2 is a flow diagram of a machine learning method for detecting malicious files.

FIG. 3 presents examples of dynamics of changes in the degree of harmfulness in relation to the number of patterns of behavior.

FIG. 4 is an example of a relationship scheme between elements of behavior patterns.

FIG. 5 is an example of a general purpose computer system, personal computer, or server.

Although the invention may have various modifications and alternative forms, the characteristic features shown by way of example in the drawings will be described in detail. It should be understood, however, that the purpose of the description is not to limit the invention to its specific embodiment. On the contrary, the purpose of the description is to cover all changes, modifications that are included in the scope of this invention, as defined by the attached formula.

Description of Embodiments

The objects and features of the present invention, methods for achieving these objects and features will become apparent by reference to exemplary embodiments. However, the present invention is not limited to the exemplary embodiments disclosed below, it can be embodied in various forms. The essence described in the description is nothing more than the specific details necessary to assist the specialist in the field of technology in a comprehensive understanding of the invention, and the present invention is defined in the scope of the attached claims.

We introduce a number of definitions and concepts that will be used in the description of embodiments of the invention.

Malicious file - a file whose execution is obviously capable of leading to unauthorized destruction, blocking, modification, copying of computer information or the neutralization of computer information protection tools.

Malicious behavior of an executable file - a set of actions that may be performed during the execution of the mentioned file and which are obviously capable of leading to unauthorized destruction, blocking, modification, copying of information or neutralization of computer information protection tools.

Malicious activity of an executable file - a set of actions performed by the said file in accordance with its malicious behavior.

The computing device of the average user is a hypothetical (theoretical) computing device that has the average characteristics of computing devices of a pre-selected group of users on which the same applications are executed as on the computing devices of the mentioned users.

A command executed by a computing device is a set of machine instructions or script instructions executed by a computing device based on parameters of said instructions, called command parameters or parameters describing said command.

Lexical analysis ("tokenization", from the English tokenizing) is the process of analytic analysis of the input sequence of characters into recognized groups (hereinafter - tokens), with the aim of forming identification sequences (hereinafter - tokens) at the output.

A token is an identification sequence formed from a token in the process of lexical analysis.

FIG. 1 is a block diagram of a machine learning system for detecting malicious files.

The structural diagram of a machine learning system consists of means for preparing training samples 111, means for generating behavior logs 112, means for generating behavior patterns 121, means for generating convolution functions 122, means for creating a detection model 131, tools for machine learning a detection model 132, means for calculating the severity 142, resource management tools 143.

In one embodiment of the system, the said machine learning system for the detection model is a client-server architecture in which the training sample preparation tool 111, the behavior log generation tool 112, the behavior pattern generation tool 121, the convolution function generation tool 122, the detection model creation tool 131 and machine learning tool detection models 132 work on the server side, and a tool for generating behavior patterns 121, a means of calculating the degree of severity and 142 and resource management tool 143 are client side.

For example, a client can be a user's computing devices such as a personal computer, laptop, smartphone, etc., and a server can be an anti-virus company’s computing devices, such as distributed server systems, which, among other things, are preliminarily performed collection and antivirus analysis of files, creation of antivirus entries, etc., while the machine learning system for detecting malicious files will be used to detect malicious files in the client, thereby increasing the effectiveness of anti-virus protection of said client.

In another example, only the antivirus company can act as a client or server, while the machine learning system for the malware detection model will be used for automated antivirus analysis of files and creation of antivirus entries, thereby increasing the efficiency of the antivirus company.

The training sample preparation tool 111 is intended for:

выборки по меньшей мере одного файла из базы файлов согласно заранее заданным правилам формирования обучающей выборки файлов, впоследствии на основании анализа выбранных файлов средство машинного обучения модели обнаружения 132 будет выполнять обучение модели обнаружения;

fetching at least one file from the file database according to predefined rules for generating a training sample of files; subsequently, based on the analysis of the selected files, the machine tool for detecting model 132 will train the detection model;

передачи выбранных файлов средству формирования журналов поведения 112.

transferring the selected files to the behavior logger 112.

In one embodiment of the system, at least one secure file and one malicious file are stored in the file database.

For example, files of the Windows operating system can be stored in the file database as safe files, and backdoor files, applications that provide unauthorized access to data and remote control of the operating system and the computer as a whole, can be stored as malicious files. At the same time, the malware detection model trained on the mentioned files using machine learning methods will be able to detect malware with the same functionality as the above mentioned backdoors with high accuracy (the higher the more files used to train the said detection model).

In another embodiment of the system, at least:

подозрительные файлы (англ. riskware) - файлы, не являющиеся вредоносными, но способные выполнять вредоносные действия;

suspicious files (eng. riskware) - files that are not malicious, but capable of performing malicious actions;

неизвестные файлы - файлы, вредоносность которых не была определена и остается неизвестной (т.е. файлы, не являющиеся безопасными, вредоносными, подозрительными и т.д.).

unknown files - files whose maliciousness has not been determined and remains unknown (i.e. files that are not safe, malicious, suspicious, etc.).

For example, files of remote administration applications (for example, RAdmin), archiving or encryption of data (for example, WinZip), etc. can act as suspicious files in the file database.

In yet another embodiment of the system, at least files are stored in the file database:

собранные антивирусными поисковыми роботами (англ. web crawler);

collected by antivirus search robots (eng. web crawler);

переданные пользователями.

submitted by users.

Moreover, the mentioned files are analyzed by anti-virus experts, including using automatic file analysis tools, for the subsequent decision on the harmfulness of the mentioned files.

For example, a file database can store files transferred by users from their computing devices to anti-virus companies to check for malware, while the transferred files can be both safe and malicious, while the distribution between the number of safe and malicious files mentioned is close to the distribution between the number of all safe and malicious files located on the computing devices of the mentioned users (i.e. the ratio of the number of said safe to the number of mentioned These malicious files differ from the ratio of the number of all safe to the number of all malicious files located on the computing devices of the mentioned users by an amount less than the specified threshold

).

)

 Unlike files transferred by users (i.e., files subjectively suspicious), files collected by anti-virus search robots designed to search for suspicious and malicious files are more often malicious.

In another embodiment of the system’s implementation, at least one of the following conditions acts as criteria according to which files are selected from the file database:

распределение между безопасными и вредоносными файлами, выбранными из базы файлов, соответствует распределению между безопасными и вредоносными файлами, расположенными на вычислительном устройстве среднестатистического пользователя;

the distribution between safe and malicious files selected from the file database corresponds to the distribution between safe and malicious files located on the computing device of the average user;

распределение между безопасными и вредоносными файлами, выбранными из базы файлов, соответствует распределению между безопасными и вредоносными файлами, собранными с помощью антивирусных поисковых роботов;

the distribution between safe and malicious files selected from the file database corresponds to the distribution between safe and malicious files collected using anti-virus search robots;

параметры файлов, выбранных из базы файлов, соответствуют параметрам файлов, расположенных на вычислительном устройстве среднестатистического пользователя;

parameters of files selected from the database of files correspond to parameters of files located on the computing device of the average user;

количество выбранных файлов соответствует заранее заданному значению, а сами файлы выбраны случайным образом.

the number of selected files corresponds to a predetermined value, and the files themselves are selected randomly.

For example, the file database contains 100,000 files, including 40% safe files and 60% malicious files. 15,000 files are selected from the file database (15% of the total number of files stored in the file database) so that the distribution between the selected safe and malicious files corresponds to the distribution between the safe and malicious files located on the average user's computing device and is 95 to 5. For this purpose, 14250 safe files (35.63% of the total number of safe files) and 750 malicious files (1.25% of the total number of malicious files) are randomly selected from the file database.

In another example, the file database contains 1250,000 files, including 95% safe files and 5% malicious files, i.e. the distribution between safe and malicious files stored in the file database corresponds to the distribution between safe and malicious files located on the computing device of the average user. Among the mentioned files, 5,000 files are randomly selected, among which with a high probability there will be ~ 4750 safe files and ~ 250 malicious files.

In another embodiment of the system, at least:

вредоносность файла, характеризующая, является ли файл безопасным, вредоносным, потенциально опасным или поведение вычислительной системы при исполнении файла не определено и т.д.;

file harmfulness, characterizing whether the file is safe, malicious, potentially dangerous, or the behavior of the computing system during file execution is not defined, etc .;

количество команд, выполненных вычислительным устройством во время исполнения файла;

the number of instructions executed by the computing device during file execution;

размер файла;

file size;

приложения, использующие файл.

applications using the file.

For example, malicious files are selected from the file database, which are scripts in the ActionScript language executed by the Adobe Flash application and do not exceed 5KB.

In another embodiment of the system, an additional means for preparing training samples 111 is intended for:

выборки по меньшей мере еще одного файла из базы файлов согласно заранее заданным правилам формирования тестовой выборки файлов, при этом впоследствии на основании анализа выбранных файлов средство машинного обучения модели обнаружения 132 будет выполнять проверку обученной модели обнаружения;

selecting at least one more file from the file database according to predefined rules for generating a test sample of files, and subsequently, based on the analysis of the selected files, the machine tool for detecting model 132 will test the trained model for detection;

передачи выбранных файлов средству формирования журналов поведения 112.

transferring the selected files to the behavior logger 112.

For example, the file database contains 75,000 files, including 20% safe files and 80% malicious files. Initially, 12,500 files are selected from the file database, among which 30% of safe files and 70% of malicious files, and subsequently, based on the analysis of the selected files, the detection model machine learning tool 132 will train the detection model, then 2,500 files are selected from the remaining 62500 files, among of which 60% of safe files and 40% of malicious files, and subsequently, based on the analysis of the selected files, the machine learning tool for detection model 132 will check the trained detection model. Data generated in the manner described above is called a cross-validation set of data.

Means of generating behavior logs 112 is intended for:

перехвата по меньшей мере одной исполняемой команды по меньшей мере во время:

intercepting at least one executable command at least during:

исполнения полученного файла,

execution of the received file,

эмуляции исполнения полученного файла, при этом эмуляция исполнения файла включает в том числе открытие упомянутого файла (например, открытие сценария интерпретатором);

emulating the execution of the received file, while emulating the execution of the file includes, among other things, opening the mentioned file (for example, opening the script by the interpreter);

определения для каждой перехваченной команды по меньшей мере одного параметра, описывающего упомянутую команду;

determining, for each intercepted command, at least one parameter describing said command;

формирования на основании перехваченных команд и определенных параметров журнала поведения полученного файла, при этом журнал поведения представляет собой совокупность перехваченных команд (далее - команда) из файла, где каждой команде соответствует по меньшей мере один определенный параметр, описывающий упомянутую команду (далее - параметр).

forming, based on the intercepted commands and certain parameters of the behavior log of the received file, the behavior log is a collection of intercepted commands (hereinafter referred to as the command) from the file, where at least one specific parameter corresponding to the mentioned command (hereinafter referred to as the parameter) corresponds to each command.

For example, commands intercepted during execution of a malicious file that collects passwords and sends them over a computer network may look like:

CreateFile, 'c: \ windows \ system32 \ data.pass'

ReadFile, 0x14ea25f7, 0xf000

connect, http://stealpass.com

send, 0x14ea25f7, 0xf000

In one embodiment of the system, the interception of commands from a file is carried out using at least:

специализированного драйвера;

specialized driver;

средства отладки (англ. debugger);

debugging tools;

гипервизора (англ. hypervisor).

hypervisor.

For example, interception of commands during the execution of a file and determination of their parameters is carried out using a driver that uses splicing (English entry) of the entry point (English entry point) of a WinAPI function.

In another example, the interception of commands when emulating a file is done directly by means of an emulator that performs the above emulation, which determines the parameters of the command that you want to emulate.

In another example, the interception of commands when executing a file on a virtual machine is carried out using the hypervisor, which determines the parameters of the command that you want to emulate.

In another embodiment of the system, at least:

API-функции;

API functions

совокупности машинных инструкций, описывающих заранее заданный набор действий (макрокоманд).

a set of machine instructions that describe a predefined set of actions (macros).

For example, very often malicious programs search for some files and modify their attributes, for which a sequence of commands is performed, such as:

FindFirstFile, 'c: \ windows \ system32 \ *. Pass', 0x40afb86a

SetFileAttributes, 'with: \ windows \ system32 \ data.pass'

FindNextFile, 0x40afb86a

CloseHandle, 0x40afb86a,

 which can in turn be described by only one command

_change_attributes, 'with: \ windows \ system32 \ *. pass'

In another embodiment of the system, each team is assigned its own unique identifier.

For example, all WinAPI functions can be assigned numbers in the range from 0x0000 to 0x8000, while each WinAPI function has its own unique number (for example, ReadFile → 0x00f0, ReadFileEx → 0x00f1, connect → 0x03A2).

In another embodiment of the system, several teams describing similar actions are assigned a single identifier.

For example, all commands, such as ReadFile, ReadFileEx, ifstream, getline, getchar, etc. that describe reading data from a file are assigned the identifier _read_data_file (0x70F0).

Means of creating patterns of behavior 121 is intended for:

формирования по меньшей мере одного шаблона поведения на основании команд и параметров, выбранных из журнала поведения, при этом журнал поведения представляет собой совокупность исполняемых команд (далее - команда) из файла, где каждой команде соответствует по меньшей мере один параметр, описывающий упомянутую команду (далее - параметр), шаблон поведения представляет собой набор из по меньшей мере одной команды и такого параметра, который описывает все команды из упомянутого набора (далее - элементы шаблона поведения);

the formation of at least one behavior template based on commands and parameters selected from the behavior log, and the behavior log is a set of executable commands (hereinafter referred to as the command) from a file, where at least one parameter describing the said command corresponds to each command (hereinafter - parameter), the behavior template is a set of at least one command and such a parameter that describes all the commands from the said set (hereinafter referred to as elements of the behavior template);

передачи сформированных шаблонов поведения средству формирования функций свертки 122;

transmitting the generated patterns of behavior to the means of forming convolution functions 122;

For example, from the behavior log, the following commands c _i (from the English command) and parameters p _i (from the English parameter) are selected:

{c ₁ , p ₁ , p ₂ , p ₃ },

{c ₂ , p ₁ , p ₄ },

{c ₃ , p ₅ },

{c ₂ , p ₅ },

{c ₁ , p ₅ , p ₆ },

{c ₃ , p ₂ }.

Based on the selected commands and parameters, behavior patterns are formed containing one command and one parameter describing the command:

{c ₁ , p ₁ }, {c ₁ , p ₂ }, {c ₁ , p ₃ }, {c ₁ , p ₅ }, {c ₁ , p ₆ },

{c ₂ , p ₁ }, {c ₂ , p ₄ }, {c ₂ , p ₅ },

{c ₃ , p ₂ }, {c ₃ , p ₅ }.

Then, based on the generated patterns, behavior patterns are additionally formed, containing one parameter and all the commands described by the mentioned parameter:

{c ₁ , c ₂ , p ₁ },

{s ₁ , s ₃ , p ₂ },

{c ₁ , s ₂ , s ₃ , p ₅ },

Then, based on the generated patterns, behavior patterns are additionally formed, containing several parameters and all the commands simultaneously described by the mentioned parameters:

{c ₁ , c ₂ , p ₂ , p ₅ }.

In one embodiment of the system implementation, the commands and parameters from the behavior log are selected based on the rules by which at least they select:

последовательно каждую i-ую команду и описывающие ее параметры, при этом шаг i является заранее заданным;

sequentially each i-th command and parameters describing it, while step i is predetermined;

команды, выполненные через заранее заданный промежуток времени (например, каждую десятую секунду) после предыдущей выбранной команды, и описывающие их параметры;

commands executed after a predetermined period of time (for example, every tenth second) after the previous selected command, and describing their parameters;

команды и описывающие их параметры, выполненные в заранее заданном временном диапазоне с начала исполнения файла;

commands and parameters describing them, executed in a predetermined time range from the beginning of file execution;

команды из заранее заданного списка и описывающие их параметры;

commands from a predefined list and describing their parameters;

параметры из заранее заданного списка и описываемые упомянутыми параметрами команды;

parameters from a predefined list and described by the command parameters mentioned;

первые или случайные k параметров команд, в случае, когда количество параметров команд больше заранее заданного порогового значения.

first or random k command parameters, in the case when the number of command parameters is greater than a predetermined threshold value.

For example, from the behavior log, all commands for working with the hard drive (such as CreateFile, ReadFile, WriteFile, DeleteFile, GetFileAttribute, etc.) and all parameters describing the selected commands are selected.

In another example, every thousandth command and all parameters describing the selected commands are selected from the behavior log.

In one embodiment of the system, the behavior logs are pre-generated from at least two files, one of which is a safe file and the other is a malicious file.

In yet another embodiment of the system, each element of the behavior pattern is associated with a characteristic such as the type of element of the behavior pattern. The type of element of the behavior template (command or parameter) is at least:

в случае, если элемент шаблона поведения может быть выражен в виде числа - «численный диапазон»

in case the element of the behavior pattern can be expressed as a number - “numerical range”

for example, for a behavior template element, which is the port _html = 80 parameter of the connect command, the type of the mentioned behavior template element can be “a numerical value from 0x0000 to 0xFFFF”,

в случае, если элемент шаблона поведения может быть выражен в виде строки - «строка»

in case the element of the behavior template can be expressed as a string - “string”

for example, for a behavior template element that is a connect command, the type of the mentioned behavior template element may be “a string of less than 32 characters”,

в случае, если элемент шаблона поведения может быть выражен в виде данных, описываемых заранее заданной структурой данных, тип упомянутого элемента шаблона поведения может быть «структура данных»

in case the element of the behavior template can be expressed in the form of data described by a predetermined data structure, the type of said element of the behavior template can be “data structure”

for example, for a behavior pattern element representing the src = 0x336b9a480d490982cdd93e2e49fdeca7 parameter of the find_record command, the type of the behavior pattern element mentioned may be “MD5 data structure”.

In yet another embodiment of the system, tokens formed on the basis of the lexical analysis of the mentioned elements of the behavior template using at least:

заранее заданных правил формирования лексем,

predefined rules for the formation of tokens,

заранее обученной рекурсивной нейронной сети (англ. recurrent neural network).

a pre-trained recursive neural network.

For example, using lexical analysis of a parameter

'with: \ windows \ system32 \ data.pass'

based on the rules for the formation of tokens:

если строка содержит путь к файлу, определить диск, на котором расположен файл;

if the line contains the path to the file, determine the drive on which the file is located;

если строка содержит путь к файлу, определить папки, в которых расположен файл;

if the line contains the path to the file, determine the folders in which the file is located;

если строка содержит путь к файлу, определить расширение файла;

if the line contains the path to the file, determine the file extension;

where the tokens are:

пути к файлу;

file paths;

папки, в которых расположены файлы;

folders in which the files are located;

имена файлов;

file names;

расширения файлов;

file extensions;

Tokens can be formed:

“File paths” →

'from:\',

“Folders in which files are located” →

'windows',

'system32',

'windows \ system32',

"File extensions" →

'.pass'.

In another example, using lexical analysis of parameters

'81 .19.82.8 ', '81 .19.72.38', '81 .19.14.32 '

based on the rule of the formation of the token:

если параметры представляют собой IP-адреса, определить битовую маску (или ее аналог, выраженный через метасимволы), описывающую упомянутые IP-адреса (т.е. такую битовую маску М, для которой верно равенство М∧IP=const для всех упомянутых IP);

if the parameters are IP addresses, define a bit mask (or its analog expressed in metacharacters) that describes the mentioned IP addresses (i.e. such a bit mask M for which the equality М∧IP = const for all the mentioned IPs is true) ;

a token can be formed:

'81 .19. *. * '.

In another example, of all the available parameters, which are numbers, form tokens of numbers in predetermined ranges:

23, 16, 7224, 6125152186, 512, 2662162, 363627632, 737382, 52,

2625, 3732, 812, 3671, 80, 3200

sort by ranges of numbers:

from 0 to 999

→ {16, 23, 52, 80, 512, 812},

from 1000 to 9999

→ {2625, 3200, 3671, 7224},

from 10000

→ {737382, 2662162, 363627632, 6125152186}

In another embodiment of the system implementation, tokens are formed from elements of the behavior template, which are strings.

For example, a behavior pattern is a path to a file containing the names of the drive, directory, file, file extension, etc. In this case, the token may be the name of the disk and the file extension.

C: \ Windows \ System32 \ drivers \ acpi.sys

→

FROM:\

* .sys

Convolution function forming means 122 is intended for:

формирования такой функции свертки от шаблона поведения, что обратная функция свертки от результата такой функции свертки над полученным шаблоном поведения будет иметь степень схожести с полученным шаблоном поведения больше заданного значения, т.е.

the formation of such a convolution function from the behavior pattern that the inverse convolution function of the result of such a convolution function over the obtained behavior pattern will have a degree of similarity with the obtained behavior pattern is greater than the specified value, i.e.

r ~ g ^-1 (g (r))

Where:

r _i - pattern of behavior,

g is the convolution function,

g ^-1 is the inverse convolution function.

передачи сформированной функции свертки средству машинного обучения модели обнаружения 132.

transmitting the generated convolution function to the machine learning detection model 132.

In one embodiment of the system, the means of forming convolution functions is additionally intended for:

вычисления признакового описания (англ. feature vector) шаблона поведения на основании полученного шаблона поведения, при этом признаковое описание шаблона поведения может быть выражено как сумма хэш-сумм от элементов шаблона поведения;

computing a feature description of the behavior pattern based on the resulting behavior pattern, wherein the feature description of the behavior pattern can be expressed as the sum of the hash sums from the elements of the behavior pattern;

формирования функции свертки от признакового описания шаблона поведения, при этом функция свертки представляет собой хэш-функцию, такую, что степень схожести вычисленного признакового описания и результата обратной хэш-функции от результата упомянутой хэш-функции от вычисленного признакового описания больше заранее заданного значения.

the formation of the convolution function from the characteristic description of the behavior pattern, wherein the convolution function is a hash function such that the degree of similarity of the calculated characteristic description and the result of the inverse hash function from the result of the said hash function from the calculated characteristic description is greater than a predetermined value.

In another embodiment of the system’s implementation, the convolution function is formed by the method of metric learning (English metric learning), i.e. so that the distance between convolutions obtained using the convolution function for behavior patterns having a degree of similarity greater than a predetermined threshold value was less than a predetermined threshold value, and for behavior patterns having a similarity degree less than a predetermined threshold value is greater than a predetermined threshold value .

For example, a characteristic description of a behavior pattern can be computed as follows:

предварительно создают пустой битовый вектор, состоящий из 100000 элементов (где для каждого элемента вектора зарезервирован один бит информации);

pre-creating an empty bit vector consisting of 100,000 elements (where for each element of the vector one bit of information is reserved);

для хранения данных о командах c_i из шаблона поведения r отводят 1000 элементов, оставшиеся 99000 элементов отводят для параметров c_i из шаблона поведения r, при этом для строковых параметров отводят 50000 элементов (с 1001 элемента по 51000 элемент), для численных - 25000 элементов (с 51001 элемента по 76000 элемент);

to store data on commands c _i from the behavior template r, 1,000 elements are allocated, the remaining 99,000 elements are allocated for parameters c _i from the behavior template r, while 50,000 elements (from 1001 elements to 51,000 elements) are allocated for string parameters, 25,000 elements for numerical (from 51001 elements to 76000 elements);

каждой команде c_i из шаблона поведения r ставят в соответствие некоторое число x_i от 0 до 999, и устанавливают соответствующий бит в созданном векторе

each team c _i from the behavior pattern r is assigned a number x _i from 0 to 999, and the corresponding bit in the created vector is set

ν [x _i ] = true;

для каждого параметра p_i из шаблона поведения r вычисляют хэш-сумму по формуле:

for each parameter p _i from the behavior pattern r, the hash sum is calculated by the formula:

для строк: y_i=1001+crc32(p_i) (mod 50000)

for strings: y _i = 1001 + crc32 (p _i ) (mod 50000)

для чисел: y_i=51001+crc32(p_i) (mod 25000)

for numbers: y _i = 51001 + crc32 (p _i ) (mod 25000)

для остального: y_i=76001+crc32(p_i) (mod 24000),

for the rest: y _i = 76001 + crc32 (p _i ) (mod 24000),

and depending on the calculated hash, set the corresponding bit in the generated vector

ν [y _i ] = true;

The described bit vector with installed elements is an indicative description of the behavior pattern r.

In another embodiment of the system, a characteristic description of the behavior pattern is calculated according to the formula:

Where:

b - the base of the positional number system (for example, for a binary vector b = 2, for a vector representing a string, i.e. a collection of characters, b = 8),

r _i is the i-th element of the behavior pattern,

h is a hash function, with 0≤h (r _i ) <b.

For example, a characteristic description of a behavior pattern can be computed as follows:

предварительно создают еще один (отличный от предыдущего примера) пустой битовый вектор, состоящий из 1000 элементов (где для каждого элемента вектора зарезервирован один бит информации);

pre-create another (different from the previous example) empty bit vector consisting of 1000 elements (where for each element of the vector one bit of information is reserved);

для каждого элемента шаблона r_i из шаблона поведения r вычисляют хэш-сумма по формуле:

for each element of the template r _i from the behavior template r calculate the hash sum by the formula:

,

,

 and depending on the calculated hash, set the corresponding bit in the generated vector

v [x _i ] = true;

In another embodiment of the system implementation, the characteristic description of the behavior pattern is a Bloom filter.

For example, a characteristic description of a behavior pattern can be computed as follows:

предварительно создают еще один (отличный от предыдущих примеров) пустой вектор, состоящий из 100000 элементов;

pre-create another (different from the previous examples) empty vector consisting of 100,000 elements;

для каждого элемента шаблона r_i из шаблона поведения r вычисляют по меньшей мере две хэш-суммы с помощью набора хэш-функций {h_j} по формуле:

for each element of the template r _i from the behavior template r, at least two hash sums are calculated using a set of hash functions {h _j } according to the formula:

x _ij = h _j (r _i )

Where:

h _j (r _i ) = crc32 (r _i ),

h _j (0) = const _j ,

and depending on the calculated hash sums, the corresponding elements in the created vector are set

ν [x _ij ] = true.

In another embodiment of the system, the size of the result of the generated convolution function from the characteristic description of the behavior template is less than the size of the mentioned characteristic description of the behavior template.

For example, the characteristic description is a bit vector containing 100,000 elements, and thus has a size of 12,500 bytes, and the result of the convolution function from the said characteristic description is a set of 8 MD5 hash sums and thus has a size of 256 bytes, i.e. ~ 2% of the size of the feature description.

In another embodiment of the system, the degree of similarity of the feature description and the result of the inverse hash function from the result of the said hash function from the calculated feature description is a numerical value in the range from 0 to 1 and is calculated according to the formula:

{h (r _i )} ∧ {g _i } ∀ {h (r _i )} = {g _i }

Where:

h (r _i ) ∧ g _i denotes the coincidence of h (r _i ) with g _i

and

{h (r _i )} - the set of results of hash functions from the elements of the behavior template,

{g _i } - the set of results of the inverse hash function of the result of the hash function of the elements of the behavior template,

r _i is the i-th element of the behavior pattern,

h is a hash function

w is the degree of similarity.

For example, the computed feature description is a bit vector

,

,

the result of the convolution function from said feature description

,

,

and the result of the inverse convolution function from the result obtained above

(where elements other than the characteristic description are marked in bold). Thus, the similarity of the feature description and the result of the inverse convolution function is 0.92.

In another embodiment of the system, said hash function, which takes a behavior pattern element as a parameter, depends on the type of behavior pattern element:

.

.

For example, to calculate the hash sum of a parameter from a behavior template that is a string containing the path to a file, the CRC32 hash function is used, any other line is the Huffman algorithm, the data set is the MD5 hash function.

In yet another embodiment of the system, the convolution function is generated from the characteristic description of the behavior template through an auto-encoder (English autoencoder), while the elements of the mentioned characteristic description of the behavior template act as input data, and data having a similarity coefficient input above a predetermined threshold value.

Detector Model Generator 131 is designed to:

создания модели обнаружения вредоносных файлов, которое включает в себя по меньшей мере:

creating a malware detection model that includes at least:

выбор метода машинного обучения модели обнаружения;

the choice of machine learning detection model;

инициализации параметров модели обучения, при этом параметры модели обучения, проинициализированные до начала машинного обучения модели обнаружения, называются гиперпарамерами (англ. hyperparameter);

initialization of the parameters of the learning model, while the parameters of the learning model, initialized before the start of machine learning of the detection model, are called hyperparameters;

depending on the file parameters selected by the training sample preparation tool 111;

передачи созданной модели обучения средству машинного обучения модели обнаружения 132.

transferring the created learning model to the machine learning detection model 132.

For example, when choosing a machine learning method for a detection model, the decision is first made whether to use an artificial neural network or random forests as the detection model, then, in the case of choosing a random forest, a separating criterion for nodes of a random forest is selected; or in the case of choosing an artificial neural network, the method of numerical optimization of the parameters of the artificial neural network is selected. At the same time, the decision to choose one or another machine learning method is made based on the effectiveness of the mentioned method when detecting malicious files (i.e. the number of errors of the first and second kind that occur when malicious files are detected) using input data (behavior patterns) of a predetermined type (i.e. data structures, the number of elements of behavior patterns, the performance of the computing device on which the malicious files, available resources of the computing device are searched etc.).

In yet another example, a machine learning method of detection model is selected based on at least:

перекрестной проверки, скользящего контроля, кросс-валидации (англ. cross-validation, CV);

cross-validation, rolling control, cross-validation (English cross-validation, CV);

математического обоснования критериев AIC, BIC и т.д.;

mathematical justification of the criteria AIC, BIC, etc .;

А/В тестирования (англ. А/В testing, split testing);

A / B testing (Eng. A / B testing, split testing);

стекинга.

stacking.

In another example, in the case of low performance computing devices random forests are chosen, otherwise, an artificial neural network.

In one embodiment of the system, machine learning of a previously created untrained detection model is performed (i.e., a detection model for which the parameters of the said model do not allow obtaining, based on the analysis of the input data, the output data with an accuracy higher than a predetermined threshold value).

In yet another embodiment of the system, at least:

градиентный бустинг на деревьях принятия решений (англ. decision-tree-based gradient boosting);

gradient boosting on decision trees (English decision-tree-based gradient boosting);

решающие деревья (англ. decision trees);

decision trees (English decision trees);

метод ближайших соседей kNN (англ. K-nearest neighbor method);

closest neighbor method kNN (English K-nearest neighbor method);

метод опорных векторов (англ. support vector machine, SVM).

support vector machine (eng. support vector machine, SVM).

In another embodiment of the system, the detection model creation tool 131 is additionally designed to create a detection model upon request from the machine learning tool 132, while certain hyperparameters and machine learning methods are selected different from the hyperparameters and machine learning methods selected for the previous detection model.

Detection model machine learning tool 132 is intended to teach a detection model in which the parameters of the detection model are calculated using the obtained convolution function over the obtained behavior patterns, where the detection model is a set of rules for calculating the degree of maliciousness of a file based on at least one behavior pattern using calculated parameters of said detection model.

For example, the detection model is trained on a well-known set of files selected by the training sample preparation tool 111, while said set of files contains 60% of safe files and 40% of malicious files.

In one embodiment of the system, the degree of maliciousness of the file is a numerical value from 0 to 1, while 0 means that the file is safe, 1 means malicious.

In yet another embodiment of the system, a method for training a detection model is selected that provides a monotonic change in the degree of harmfulness of a file depending on a change in the number of behavior patterns generated based on an analysis of the behavior log.

For example, a monotonic change in the severity of a file leads to the fact that when analyzing each subsequent behavior pattern, the calculated severity is no less than the previously calculated severity (for example, after analyzing the 10th behavior pattern, the calculated severity is 0.2, after analysis 50th pattern of behavior - 0.4, and after analysis of the 100th pattern of behavior - 0.7).

In another embodiment of the system, an additional machine learning tool for the detection model 132 is intended for:

выполнения проверки обученной модели обнаружения на полученных журналах поведения, сформированных на основании анализа файлов из тестовой выборки файлов, с целью определить корректность определения вредоносности файлов из тестовой выборки файлов;

checking the trained detection model on the received behavior logs generated based on the analysis of files from the test file selection in order to determine the correctness of determining the harmfulness of files from the test file selection;

в случае отрицательного результата проверки передачи запроса по меньшей мере:

in case of a negative result of checking the transfer of the request at least:

средству подготовки обучающих выборок 111 для подготовки выборки файлов, отличной от текущей, на которой производилось обучение модели обнаружения;

the training sample preparation tool 111 for preparing a sample of files other than the current one on which the detection model was trained;

средству создания модели обнаружения 131 для создания новой модели обнаружения, отличной от текущей.

detection model tool 131 to create a new detection model different from the current one.

The verification of the trained detection model is as follows. The mentioned detection model was trained on the basis of a set of 111 files selected by the training sample preparation tool for which it was known whether they were safe or malicious. In order to verify that the malware detection model has been trained correctly, i.e. said detection model will be able to detect malicious files and skip safe files, the said model is checked. To this end, using the aforementioned detection model, it is determined whether the files from another set of 111 files selected by the training sample preparation tool are malicious, while it is known whether the said files are malicious. Thus, it determines how many malicious files were “skipped” and how many safe files were found. If the number of skipped malicious and detected safe files is greater than a predetermined threshold value, then the mentioned detection model is considered to be incorrectly trained and it is necessary to perform its machine learning again (for example, on another training sample of files using different detection model parameters, etc. .).

For example, when testing a trained model, the number of errors of the first and second kind when detecting malicious files from a test sample of files is checked. If the number of the mentioned errors exceeds a predetermined threshold value, a new training and test selection of files is selected and a new detection model is created.

In another example, the training sample of files contained 10,000 files, of which 8,500 were malicious, and 1,500 were safe. After the detection model was trained, it was tested on a test sample of files containing 1200 files, of which 350 were malicious, and 850 were safe. According to the results of the scan, 15 out of 350 malicious files could not be detected (4%), while at the same time, 102 out of 850 safe files (12%) were mistakenly recognized as malicious. In the case when the number of undetected malicious files exceeds 5% or accidentally detected safe files exceeds 0.1%, the trained detection model is recognized as incorrectly trained.

In one embodiment of the system, an additional system behavior log is generated based on a previously generated system behavior log and commands intercepted after the formation of the system behavior log.

For example, after the start of the execution of a file for which a verdict on the harmfulness or security of the said file is to be issued, intercepted executable commands and parameters describing them are recorded in the behavior log. Based on the analysis of the mentioned commands and parameters, the harmfulness coefficient of the mentioned file is calculated. If, according to the results of the analysis, a decision to recognize the file as malicious or safe was not made, command interception can continue. Intercepted commands and parameters describing them are added to the old behavior log or to the new behavior log. In the first case, the harmfulness coefficient is calculated based on an analysis of all the command and parameters recorded in the behavior log, i.e. and those previously used to calculate the malware rate.

The malware severity calculation tool 142 is intended for:

вычисления степени вредоносности на основании полученного от средства формирования журнала поведения 142 журнала поведения и от средства машинного обучения модели обнаружения 132 модели обнаружения, при этом степень вредоносности файла представляет собой количественную характеристику (например, лежащую в диапазоне от 0 - файл обладает исключительно безопасным поведением до 1 - упомянутый файл обладает заранее заданным вредоносным поведением), описывающую вредоносное поведение исполняемого файла;

calculating the degree of severity based on the behavior log received from the means of generating the behavior log 142 of the behavior log and from the machine learning tool of the detection model 132 of the detection model, while the degree of severity of the file is a quantitative characteristic (for example, lying in the range from 0 - the file has exceptionally safe behavior up to 1 - the said file has a predefined malicious behavior) that describes the malicious behavior of the executable file;

передачи вычисленного степени вредоносности средству управления ресурсами 143.

transmitting the calculated severity to the resource manager 143.

The resource management tool 143 is intended for analyzing the degree of severity of the allocation of computing resources of a computer system for use in securing a computer system.

In one embodiment of the system, at least:

объем свободной оперативной памяти;

amount of free RAM;

объем свободного места на жестких дисках;

amount of free space on hard drives;

свободное процессорное время (кванты процессорного времени), которое может быть потрачено на антивирусную проверку (например, с большей глубиной эмуляции).

free processor time (quanta of processor time) that can be spent on anti-virus scanning (for example, with greater emulation depth).

In another embodiment of the system, the analysis of the degree of harmfulness consists in determining the dynamics of the change in the value of the degree of harmfulness after each of the previous calculations of the degree of harmfulness and in the case of at least:

увеличение значения степени вредоносности - выделения дополнительных ресурсов компьютерной системы;

increasing the value of the degree of severity - allocation of additional resources of a computer system;

уменьшение значения степени вредоносности - освобождение ранее выделенных ресурсов компьютерной системы.

reducing the value of the degree of severity - the release of previously allocated resources of a computer system.

FIG. 2 is a flow diagram of a machine learning method for detecting malicious files.

The block diagram of the machine learning method for detecting malicious files contains step 211, which prepares training samples of files, step 212, which generates behavior logs, step 221, which generates behavior patterns, step 222, which generates convolution functions, step 231, which creates a detection model, step 232, where I train the detection model, step 241, which tracks the behavior of the computer system, step 242, which calculates the degree of harmfulness, step 243, which controls the resource s computer system.

At step 211, using the training sample preparation means 111, at least one file is selected from the file base according to predetermined criteria, and the detection model will be trained based on the selected files at step 232.

At step 212 using the means of generating behavior logs 112:

перехватывают по меньшей мере одну команду по меньшей мере во время:

intercept at least one command at least during:

исполнения выбранного на этапе 211 файла,

executing the file selected in step 211,

эмуляции работы выбранного на этапе 211 файла;

emulating the operation of the file selected in step 211;

определяют для каждой перехваченной команды по меньшей мере один параметр, описывающий упомянутую команду;

determining for each intercepted command at least one parameter describing said command;

формируют на основании перехваченных команд и определенных параметров журнал поведения полученного файла, при этом журнал поведения представляет собой совокупность перехваченных команд (далее - команда) из файла, где каждой команде соответствует по меньшей мере один определенный параметр, описывающий упомянутую команду (далее - параметр).

form, based on the intercepted commands and certain parameters, the behavior log of the received file, while the behavior log is a collection of intercepted commands (hereinafter referred to as the command) from the file, where at least one specific parameter corresponding to the mentioned command (hereinafter referred to as the parameter) corresponds to each command.

At step 221, using the means of generating behavior patterns 121, at least one behavior pattern is generated based on commands and parameters selected from the behavior log generated in step 212, the behavior log being a set of executable commands (hereinafter referred to as the command) from a file, where each command corresponds to at least one parameter that describes the command (hereinafter referred to as the parameter), the behavior pattern is a set of at least one command and such a parameter that describes everything teams from the mentioned set.

At step 222, using the means of generating convolution functions 122, such a convolution function is generated from the behavior pattern generated in step 221 that the inverse convolution function from the result of such a convolution function over said behavior pattern will have a degree of similarity with said behavior pattern greater than a predetermined value.

At step 231, using the detection model creation tool 131, a detection model is created for which at least:

выбирают метод машинного обучения модели обнаружения;

choose a machine learning method for the detection model;

инициализируют параметры модели обучения, при этом параметры модели обучения, проинициализированные до начала машинного обучения модели обнаружения, называются гиперпарамерами (англ. hyperparameter),

initialize the parameters of the learning model, while the parameters of the learning model, initialized before the start of machine learning of the detection model, are called hyperparameters,

depending on the file parameters selected in step 211.

In step 232, the detection model created in step 231 is trained in the machine learning tool of the detection model 132, in which the parameters of the detection model are calculated using the convolution function generated in step 222 over the behavior patterns generated in step 221, where the detection model where the detection model is a set of rules for calculating the degree of harmfulness of a file based on at least one behavior pattern using the calculated parameters of the mentioned model about aruzheniya.

At step 241, using the behavior tracking means 141:

перехватывают по меньшей мере одну команду, исполняемую файлами, работающими в компьютерной системе;

intercepting at least one command executed by files operating in a computer system;

формируют на основании перехваченных команд журнал поведения системы.

form a system behavior log based on intercepted commands.

In step 242, the degree of severity is calculated using the severity calculation means 142 based on the system behavior log generated in step 241 and the detection model trained in step 232.

At step 243, using the resource management tool 143 based on the analysis of the severity calculated at step 242, computing resources are allocated for use in securing a computer system.

FIG. 3 presents examples of dynamics of changes in the degree of harmfulness in relation to the number of patterns of behavior.

Examples of dynamics of changes in the degree of harmfulness in terms of the number of patterns of behavior include a graph of the dynamics of an arbitrary change in the degree of harmfulness in terms of the number of patterns of behavior generated during the execution of a malicious file 311, a graph of the dynamics of a monotonous change in the degree of harmfulness in terms of the number of patterns of behavior formed during the execution of a malicious file 312, arbitrary changes in the degree of harmfulness from the number of behavior patterns generated during execution without asnogo file 321, a graph of the dynamics of a monotone change in the degree of harmfulness of the number of patterns of behavior generated during a secure file 322 execution.

In one embodiment of the system, the severity of the executable file takes a value in the range from 0 (the file has exceptionally safe behavior) to 1 (the file has a predefined malicious behavior).

The graph 311 shows the dynamics of an arbitrary change in the degree of harmfulness from the number of behavior patterns generated during the execution of the malicious file.

Initially, during the execution of the mentioned file, the number of generated patterns of behavior is small, in addition, the malicious activity of the executable file may be absent or minimal (for example, data initialization is inherent to many, including safe files), therefore, the calculated severity is slightly different from 0 and not exceeds a predetermined threshold value (hereinafter referred to as the security criterion) upon exceeding which the behavior of the executable file will no longer be considered safe (by r In the graphic, the mentioned threshold value is indicated by a dashed line).

However, over time, the malicious activity of the executable file increases and the degree of harmfulness begins to approach 1, exceeding the security criterion, while the degree of harmfulness may not reach a predetermined threshold value (hereinafter referred to as the harmfulness criterion), above which the behavior of the executable file will be considered malicious (on the graph mentioned the threshold value is indicated by a dashed line).

After a period of growth, malicious activity may cease and the degree of harmfulness will again tend to 0 (moment A). At a certain point, the degree of harmfulness becomes greater than the harmfulness criterion (moment B) and the behavior of the executable file is recognized as malicious and, as a result, the file itself is recognized as malicious.

At the same time, the moment the file is recognized as malicious can occur much later than the beginning of the growth of malicious activity, since the described approach responds well to a sharp increase in the degree of harmfulness, which occurs most often with prolonged, clearly pronounced malicious activity of the executable file.

In the case when malicious activity occurs occasionally (the left side of the graph 311), the calculated degree of harmfulness may not reach a value after which a decision is made on the harmfulness of the behavior of the executable file and, therefore, the harmfulness of the executable file itself.

In the case when the severity is not calculated on the basis of each generated behavior pattern (for example, due to the fact that the performance of the computing device is low), it is possible that the severity will be calculated at time A (when malicious activity begins) and moment C (when malicious activity ends), and will not be calculated at time В (when malicious activity occurs), because of which the calculated degrees of harmfulness will not exceed the harmfulness criterion, the activity is The executable file will not be recognized as malicious and, therefore, the malicious file will not be detected.

Charts 312 show the dynamics of a monotonous change in the degree of harmfulness in relation to the number of behavior patterns generated during the execution of a malicious file.

Initially, during the execution of the mentioned file, the number of generated patterns of behavior is small, in addition, the malicious activity of the executable file may be absent or minimal (for example, data initialization is inherent to many, including safe files), therefore, the calculated severity is slightly different from 0 and not exceeds a predetermined threshold value (hereinafter referred to as the security criterion) upon exceeding which the behavior of the executable file will no longer be considered safe (by r In the graphic, the mentioned threshold value is indicated by a dashed line).

However, over time, the malicious activity of the executable file increases and the degree of harmfulness begins to approach 1, exceeding the security criterion, while the degree of harmfulness may not reach a predetermined threshold value (hereinafter referred to as the harmfulness criterion), above which the behavior of the executable file will be considered malicious (on the graph mentioned the threshold value is indicated by a dashed line).

After a growth period (moments AB), malicious activity may cease (moments BA), but the degree of harmfulness will no longer decrease, and with any malicious activity of the executable file it will only continue to grow. At a certain point, the degree of harmfulness becomes greater than the harmfulness criterion (moment D) and the behavior of the executable file is recognized as malicious and, as a result, the file itself is recognized as malicious.

At the same time, the moment the file is recognized as malicious can occur immediately after the detection of malicious activity, since the described approach responds well to a smooth increase in the degree of harmfulness, which occurs both with prolonged, clearly pronounced malicious activity of the executable file, and with frequent episodic mild malicious activity.

In the case when malicious activity occurs sporadically (the left side of the graph 312), the calculated degree of malware can reach a value over time, after which a decision is made about the harmfulness of the behavior of the executable file and the harmfulness of the executable file itself.

In the case when the severity is not calculated on the basis of each generated behavior pattern (for example, due to the fact that the performance of the computing device is low), it is possible that the severity will be calculated at time A (when malicious activity begins) and moment C (when malicious activity ends), and will not be calculated at time В (when malicious activity occurs), however, since the degree of harmfulness changes the monotonously calculated degrees of malicious the worms only increase their values and at time C the degree of harmfulness will exceed the harmfulness criterion, the activity of the executable file will be recognized as malicious and, therefore, the malicious file will be detected.

The graph 321 shows the dynamics of an arbitrary change in the degree of harmfulness from the number of behavior patterns generated during the execution of a safe file.

Initially, when executing the mentioned file, the number of generated behavior patterns is small, in addition, malicious activity as such is absent in the executable file, but “suspicious” commands can also be executed when executing malicious files (for example, deleting files, transferring data over a computer network, etc.) .d.), therefore, the calculated degree of harmfulness differs from 0 and does not exceed a predetermined threshold value (hereinafter referred to as the security criterion) above which the behavior is executed of the file will cease to be considered safe (on the graph, the mentioned threshold value is indicated by a dashed line).

However, over time, the malicious activity of the executable file increases due to the execution of a large number of “suspicious” commands and the degree of harmfulness begins to approach 1, while the degree of harmfulness may not reach a predetermined threshold value (hereinafter referred to as the harmfulness criterion) above which the behavior of the executable file will be be considered malicious (on the graph, the mentioned threshold value is indicated by a dashed line), but may exceed the security criterion, thus the file may cease to itatsya safe and become a "suspicious".

After a period of growth, malicious activity may cease and the degree of harmfulness will again tend to 0 (moment C).

In the case when the severity is not calculated on the basis of each generated behavior pattern (for example, because the performance of the computing device is low), it is possible that the severity will be calculated at the moment B (when the activity is most similar to the malicious one, t .e. becomes “suspicious”) and will not be calculated at time A (when “suspicious” activity increases) and at time C (when “suspicious” activity decreases), which is why the calculated degree is harmful ti exceeds the safety criterion, the executable file activity will be recognized as "suspicious" (will not be recognized as safe) and thus secure the file is found to be safe.

The graph 322 shows the dynamics of a monotonous change in the degree of harmfulness from the number of behavior patterns generated during the execution of a safe file.

Initially, when executing the mentioned file, the number of generated behavior patterns is small, in addition, malicious activity as such is absent in the executable file, but “suspicious” commands can also be executed when executing malicious files (for example, deleting files, transferring data over a computer network, etc.) .d.), therefore, the calculated degree of harmfulness differs from 0 and does not exceed a predetermined threshold value (hereinafter referred to as the security criterion) above which the behavior is executed of the file will cease to be considered safe (on the graph, the mentioned threshold value is indicated by a dashed line).

However, over time, the malicious activity of the executable file increases due to the execution of a large number of “suspicious” commands and the degree of harmfulness begins to approach 1, while the degree of harmfulness may not reach a predetermined threshold value (hereinafter referred to as the harmfulness criterion) above which the behavior of the executable file will be be considered malicious (on the graph, the threshold value indicated by a dashed line), and may also not exceed the security criterion, thus the file will be sold To be considered safe.

After a growth period (moments AB), malicious activity may cease (moments BA), but the degree of harmfulness will not decrease anymore, and for any malicious activity of the executable file it will only continue to grow, but not exceed the safety factor, thus the activity of the executable file will be considered safe and, therefore, the mentioned file will be considered safe.

In the case when the severity is not calculated on the basis of each generated behavior pattern (for example, because the performance of the computing device is low), it is possible that the severity will be calculated at the moment B (when the activity is most similar to the malicious one, t .e. becomes “suspicious”) and will not be calculated at time A (when “suspicious” activity increases) and at time C (when “suspicious” activity decreases), however, since the degree is harmful STI changes monotonically calculated degree of harmfulness only increase their value in points A, B, C, degree of harmfulness does not exceed the safety criterion, the activity of the executable file is found to be safe and therefore safe file is deemed safe.

At the same time, the moment of recognition of the file as “suspicious” may not occur after the detection of “suspicious” activity, since the described approach ensures a smooth increase in the degree of harmfulness, which avoids sharp peaks when the degree of harmfulness increases.

FIG. 4 is an example of a relationship scheme between elements of behavior patterns.

An example of a relationship between elements of behavior patterns contains commands 411 (empty circles), parameters 412 (hatched circles), an example of a behavior pattern with one parameter 421, an example of a behavior pattern with one command 422.

During the execution of the file, commands 411 were intercepted and the parameters 412 describing them were determined:

CreateFile 0x24e0da54 '.dat'

{c1, p1, p2}

ReadFile 0x24e0da54 '.dat'

{c2, p1, p2}

DeleteFile 0x24e0da54 '.dat' 'with: \'

{c3, p1, p2, p3}

CreateFile 0x708a0b32 'dat' 0x3be06520

{c1, p2, p3, p5}

WriteFile 0x708a0b32

{c4, p3}

WriteFile 0x708a0b32 0x3be06520 0x9902a18d1718b5124728f9 0

{c4, p3, p5, p6, p7}

CopyMemory 0x3be06520 0x9902a18d1718b5124728f9

{c5, p4, p5, p6}

ReadFile 0x9902a18d1718b5124728f9 0

{c2, p6, p7}

Based on the mentioned commands 411 and parameters 412, behavior patterns (421, 422) are generated and relationships between the elements of the behavior patterns are determined.

At the first stage, templates are formed containing one command 411 and one parameter 412 describing the command:

{c1, p1} {c3, p2} {c1, p2} {c3, p3} {c1, p3} {c4, p3} {c1, p5} {c4, p5} {c2, p1} {c4, p6} {c2, p2} {c4, p7} {c2, p6} {c5, p4} {c2, p7} {c5, p5} {c3, p1} {c5, p6}

In the given example, on the basis of 8 intercepted teams (with parameters describing them), 19 patterns of behavior were generated.

At the second stage, templates are formed containing one parameter 412 and all commands 411 described by the mentioned parameter 412:

{c1, c2, c3, p1} {c1, c3, c4, p3} {c1, c2, c3, p2} {c5, p4} {c1, c4, c5, p5} {c2, c4, p7} {c2, c4, c5, p6}

In the given example, based on 8 intercepted commands (with the parameters describing them), 7 additional behavior patterns were generated.

At the third stage, templates are formed containing several parameters 412 and all commands 411 described by the mentioned templates 412:

{c1, c2, c3, p1, p2} {c2, c4, p6, p7} {c4, c5, p5, p6}

In the given example, on the basis of 8 intercepted teams (with parameters describing them), 3 behavior patterns were additionally generated.

FIG. 5 is an example of a general purpose computer system, a personal computer or server 20 comprising a central processor 21, a system memory 22, and a system bus 23 that contains various system components, including memory associated with the central processor 21. The system bus 23 is implemented as any prior art bus structure comprising, in turn, a bus memory or a bus memory controller, a peripheral bus and a local bus that is capable of interacting with any other bus architecture. The system memory contains read-only memory (ROM) 24, random access memory (RAM) 25. The main input / output system (BIOS) 26, contains basic procedures that ensure the transfer of information between the elements of the personal computer 20, for example, at the time of loading the operating ROM systems 24.

The personal computer 20 in turn contains a hard disk 27 for reading and writing data, a magnetic disk drive 28 for reading and writing to removable magnetic disks 29, and an optical drive 30 for reading and writing to removable optical disks 31, such as a CD-ROM, DVD -ROM and other optical information carriers. The hard disk 27, the magnetic disk drive 28, the optical drive 30 are connected to the system bus 23 through the interface of the hard disk 32, the interface of the magnetic disks 33 and the interface of the optical drive 34, respectively. Drives and associated computer storage media are non-volatile means of storing computer instructions, data structures, software modules and other data of a personal computer 20.

The present description discloses an implementation of a system that uses a hard disk 27, a removable magnetic disk 29, and a removable optical disk 31, but it should be understood that other types of computer storage media 56 that can store data in a form readable by a computer (solid state drives, flash memory cards, digital disks, random access memory (RAM), etc.) that are connected to the system bus 23 through the controller 55.

Computer 20 has a file system 36 where the recorded operating system 35 is stored, as well as additional software applications 37, other program modules 38, and program data 39. The user is able to enter commands and information into personal computer 20 via input devices (keyboard 40, keypad “ the mouse "42). Other input devices (not displayed) can be used: microphone, joystick, game console, scanner, etc. Such input devices are, as usual, connected to the computer system 20 via a serial port 46, which in turn is connected to the system bus, but can be connected in another way, for example, using a parallel port, a game port, or a universal serial bus (USB). A monitor 47 or other type of display device is also connected to the system bus 23 via an interface such as a video adapter 48. In addition to the monitor 47, the personal computer can be equipped with other peripheral output devices (not displayed), for example, speakers, a printer, etc. .

The personal computer 20 is capable of operating in a networked environment, using a network connection with another or more remote computers 49. The remote computer (or computers) 49 are the same personal computers or servers that have most or all of the elements mentioned earlier in the description of the creature the personal computer 20 of FIG. 5. Other devices, such as routers, network stations, peer-to-peer devices, or other network nodes, may also be present on the computer network.

Network connections can form a local area network (LAN) 50 and a wide area network (WAN). Such networks are used in corporate computer networks, internal networks of companies and, as a rule, have access to the Internet. In LAN or WAN networks, the personal computer 20 is connected to the local area network 50 via a network adapter or network interface 51. When using the networks, the personal computer 20 may use a modem 54 or other means of providing communication with a global computer network such as the Internet. The modem 54, which is an internal or external device, is connected to the system bus 23 via the serial port 46. It should be clarified that the network connections are only exemplary and are not required to display the exact network configuration, i.e. in reality, there are other ways to establish a technical connection between one computer and another.

In conclusion, it should be noted that the information provided in the description are examples that do not limit the scope of the present invention defined by the claims.

Claims (29)

1. The machine learning system of the malware detection model, which contains: a) a tool for generating patterns of behavior of malicious files designed to:

generating at least one behavior pattern of the malicious file based on the executable commands and parameters selected from the behavior log, the behavior log being a set of executable commands (hereinafter referred to as the command) from the file, where each command corresponds to at least one parameter describing the aforementioned a command (hereinafter referred to as a parameter), a malicious file behavior pattern is a set of at least one command and such a parameter that describes all the commands from the said set;

transferring the generated malicious file behavior patterns to the convolution function generation tool; b) means for the formation of convolution functions, intended for:

the formation of such a convolution function from the behavior pattern of the malicious file that the inverse convolution function from the result of such a convolution function over the resulting behavior pattern of the malicious file will have a degree of similarity with the obtained behavior pattern of the malicious file is greater than the specified value;

transmitting the generated convolution function to the machine learning detection model; c) a machine learning tool for the detection model, designed to:

training a detection model in which the parameters of the detection model are calculated using the obtained convolution function over the obtained behavior patterns of the malicious file, where the detection model is a set of rules for calculating the degree of maliciousness of a file based on at least one behavior pattern of the malicious file using the calculated parameters of the said detection model . 2. The system of claim 1, wherein the behavior logs are pre-generated from at least two files, one of which is a secure file and the other is a malicious file. 3. The system according to claim 1, in which machine learning of a previously created untrained detection model is performed. 4. The system of claim 1, wherein at least:

gradient boosting on decision trees;

decisive trees;

nearest neighbors method kNN;

support vector method. 5. The system according to claim 1, in which a method for training a detection model is selected that provides a monotonic change in the degree of maliciousness of a file depending on changes in the number of malicious file behavior patterns generated based on an analysis of the behavior log. 6. A method of machine learning a model for detecting malicious files, the method comprising the steps that are implemented using the tools from the system according to claim 1 and in which: a) form at least one pattern of behavior of the malicious file based on the executable commands and parameters selected from the behavior log, while the behavior log is a set of executable commands (hereinafter - the command) from the file, where at least one parameter corresponds to each command, describing the command (hereinafter referred to as the parameter), the behavior pattern of the malicious file is a set of at least one command and such a parameter that describes all the commands from the said set; b) form such a convolution function from the malicious file behavior pattern that the inverse convolution function from the result of such a convolution function over the generated malicious file behavior pattern will have a degree of similarity with the generated malicious file behavior pattern greater than the specified value; c) they teach a detection model in which the parameters of the detection model are calculated using the generated convolution function over the obtained malicious file behavior patterns, where the detection model is a set of rules for calculating the malware severity level based on at least one malicious file behavior pattern using the calculated parameters of the aforementioned detection models. 7. The method of claim 6, wherein the behavior logs are pre-generated from at least two files, one of which is a secure file and the other is a malicious file. 8. The method of claim 6, wherein machine learning of a previously created untrained detection model is performed. 9. The method of claim 6, wherein at least:

gradient boosting on decision trees;

decisive trees;

nearest neighbors method kNN;

support vector method. 10. The method according to p. 6, which selects a method for training a detection model that provides a monotonic change in the degree of harmfulness of a file depending on changes in the number of malicious file behavior patterns generated based on an analysis of the behavior log.

Images