KR102184772B1 - Automated Smart Contract Tagging System based on Tag Recommendation Model - Google Patents

Abstract

A smart contract recommendation system that analyzes a user's smart contract and recommends a tag is launched. Various learning methods to extract tags by analyzing known smart contracts are presented, and when a user inputs a smart contract that the user wants to analyze into the system of the present invention using the learned models, a tag matching the smart contract Is recommended to the user, and as a result, there is an advantage of being able to quickly and easily check the smart contract that the user can trust.

Description

Automatic Smart Contract Tagging System based on Tag Recommendation Model {Automated Smart Contract Tagging System based on Tag Recommendation Model}

The present invention relates to an automated smart contract recommendation system, and more particularly, to a method and system for automatically classifying/classifying smart contracts or recommending them to users using a tag method from information of smart contracts.

A smart contract can be defined as a computer program that automatically executes the written contract content when a pre-defined condition on Ethereum is satisfied. It is used for various purposes such as ICO (Initial Coin Offering), Ethereum token issuance, gambling site operation, etc. because it not only handles tasks that should automatically operate on the blockchain, but also processes cryptocurrency. .

Since such a smart contract can process the transmitted Ethereum, if the smart contract is abused, financial fraud problems such as ponzi and pyramid schemes may occur.

In addition, in order to check what content the smart contract is written with, there is an inconvenience of interpreting the execution code in the form of an executable file stored in the blockchain or the source code uploaded by the author. Without this, there is a problem that it cannot be interpreted at all.

As a result, in order to reduce the damage caused by the asymmetry of information between the service provider and the service user using the smart contract, it is necessary to provide information that can identify the contents of the smart contract.

Publication No. 10-2018-0068888, Method and device for purchasing game items using smart contracts Registration No. 10-1778768, IoT device control method and system

An object of the present invention to solve the above problems is to provide a system capable of recommending tags of an automated smart contract.

The present invention for achieving the above object is a first smart contract analysis unit (1100), a first smart contract analysis unit for generating a plurality of learning and evaluation datasets by extracting a plurality of parameters from a plurality of smart contracts to be learned A smart contract recommendation model learning unit 1200 that receives the training data set from and trains it using a specific learning model, and compares the evaluation data set with data learned from the smart contract recommendation model learning unit, and a smart contract tag It characterized in that it comprises a smart contract optimal model selection unit 1300 for selecting an optimal model capable of recommending.

In addition, the present invention provides a smart contract optimal model selection unit 1300 that extracts a plurality of parameters from a plurality of smart contracts and recommends an optimal tag through learning a plurality of models, and a user extracts a plurality of parameters from the smart contract. 2 Smart contract analysis unit 2100, a smart contract model application unit 2200 that applies the optimal model selected from the smart contract optimal model selection unit 1300 to parameters extracted from the second smart contract analysis unit 2100 And a tag recommendation application unit 2300 for recommending a tag selected from the smart contract model application unit 2200.

In addition, the plurality of parameters extracted from the smart contract is characterized by using an executable code (opcode) obtained by disassembling the binary code of the smart contract, and information extracted from the source code of the smart contract.

In addition, the specific learning model consists of a plurality of models, and each of the plurality of models selectively selects two or more of Deep Neural Networks, Deep Collaborative Filtering, and K-Nearest Neighbors (KNN). It is characterized in that it is used in combination.

In addition, the information extracted from the execution code (opcode) includes a length, a context, and a component, and the information extracted from the source coat includes a title (contract name) or a function name (function name) of the smart contract. It is characterized.

In the case of using the smart contract recommendation system according to the present invention as described above, a tag capable of identification can be recommended and provided to the user so that the smart contract can be easily identified or distinguished.

In addition, it provides an effect of allowing a user to intuitively know the practical use and risk of a smart contract through the tag generated according to the present invention.

In addition, the smart contract recommendation method and recommendation system of the present invention can guarantee the convenience of a smart contract user through an automated process.

In addition, the use of the present invention ensures that all users can easily and fairly use the service of the smart contract regardless of whether the users have a professional system skill to check the contents of the smart contract.

As a result, there is an advantage that users can be easily protected from various financial frauds (such as Ponzi fraud) caused by smart contracts.

1 is a conceptual diagram of an automated smart contract recommendation system using a smart contract tag recommendation model
2 is a detailed configuration diagram of an automated smart contract recommendation system using a smart contract tag recommendation model.
3 is a conceptual diagram of a deep collaborative filtering model.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, but one or more other features. It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof, does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Does not.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 or 2 illustrates a conceptual and detailed configuration diagram of an automated smart contract recommendation system 10000 using a smart contract tag recommendation model.

The smart contract recommendation system 10000 of the present invention may be implemented in a form in which a user requests a recommendation for a smart contract through the user device 700 as shown in FIG. 1. That is, when a user wants to obtain information related to a smart contract, the user requests a smart contract that is the object of which he/she wants to know information to a recommendation system such as the present invention online/offline, and the recommendation system provides the user with a tag suitable for the requested smart contract. I will recommend it.

In addition, the smart contract recommendation system 10000 includes a smart contract tag recommendation model build-up unit 1000 and a system of the present invention for the purpose of building the system of the present invention to efficiently extract tags from the smart contract. It consists of a smart contract tag recommendation unit (2000) for the purpose of applying (system application process).

The smart contract tag recommendation model build-up unit 1000 constitutes the upper part of FIG. 2, and the smart contract tag recommendation unit 2000 constitutes the lower part of FIG. 2.

The smart contract tag recommendation model build-up unit 1000 may again include a first smart contract analysis unit 1100, a smart contract recommendation model learning unit 1200, and a smart contract optimal model selection unit 1300.

Specifically, the first smart contract analysis unit 1100 plays a role of generating a plurality of learning and evaluation datasets by extracting a plurality of parameters from a plurality of smart contracts to be learned.

More specifically, by obtaining multiple smart contracts that can obtain source code and binary code from the blockchain system, disassemble the binary code to obtain an execution code (Opcode), and extract component, context information, and length information from it do. After concatenating the extracted information, make a dataset.

Here, in relation to the components, the execution code (opcode) can be thought of as a detailed action that must be performed in order to implement the entire contract. You may have to take every action you can (opcode) to fulfill a specific purpose, but some actions can achieve that goal. In other words, because the types of actions required for different purposes are different, we define and extract the types of opcodes required to implement the contract as components of the smart contract.

Specifically, the components of the above-described execution code are one-hot encoding to obtain information on what actions the smart contract achieves its purpose with, so that the executed action is 1 and the unexecuted action is 0. One-hot encoding) can be implemented by converting the components of a smart contract into a 66-dimensional vector format.

Here, in relation to context information, even if the smart contract has the same component, different results may be produced according to the execution order of the execution code (opcode). In order to compensate for the limitation of such component information, context information of an executable code (opcode) is extracted using a document embedding technique.

Specifically, the context information of the execution code described above can be obtained by converting it into a vector form through the Doc2Vec algorithm. The Doc2Vec algorithm is a technique that analyzes the context by considering the order of words contained in a document and converts it into a vector form. By applying this, context information is derived from the documents listed in the order of execution of the opcode of the smart contract and converted into a real value vector. The dimension of the vector can be arbitrarily set, and a 32-dimensional vector is used as an example.

Here, the length information means the length of the execution code, and the length of the execution code is usually determined according to the amount of contract contents using the smart contract. Depending on the purpose of the smart contract, there are sub-contract contents that should be basically created, and the more complex the contents, the longer the execution code is.

Specifically, the length information of the execution code described above is a two-dimensional consisting of integer values using the string length of the binary code (bytecode) and the length of the execution code (opcode) document representing the action record of the smart contract (the number of actions to be executed). It can be converted to a vector and used.

In addition, the title and function name of the smart contract are extracted from the source code, tag information is generated, and then added to the above-described data set.

Here, in the case of a tag generated from the source code, specifically, a contract name given by the creator of the smart contract and a function name included in the source code uploaded by the creator are extracted. At this time, the fact that the function name normally used by the author can reflect the unique function or characteristic of the smart contract, and the probability that functions of the same type are described in the same function name in the source code of the smart contract used for financial fraud. In that it is high, it is meaningful to reflect the function name in tag recommendation.

More specifically, to generate the tag information, first, among the smart contracts stored in the blockchain, n samples of smart contracts including title information and source code information are extracted. In order to obtain the key words used in the title and function names of the extracted samples, unnecessary numbers and special characters are first removed, and words are created using regular expressions based on spaces and capital letters. The generated word contains abbreviated expressions, typos, and unique words, so it will be implemented by proceeding with the process of unifying abbreviated expressions, correcting typos, and removing unique words through total investigation, and selecting m words with representativeness. I can.

As a result, the collected data set is divided into a training data set for learning and an evaluation data set for evaluation, among which the training data set is used as training data in the smart contract recommendation model learning unit 1200 to be described later. In addition, the above-described evaluation dataset is used to evaluate the results from the trained model later.

In addition, the smart contract recommendation model learning unit 1200 receives the training data set from the first smart contract analysis unit and performs a role of learning using a specific learning model.

At this time, the specific learning model is composed of a plurality of models, and each of the plurality of models selectively selects at least two of Deep Neural Networks, Deep Collaborative Filtering, and K-Nearest Neighbors (KNN). Can be used in combination.

The above-described plurality of models will be described in detail in the description of FIG. 3 to be described later.

In addition, the smart contract optimal model selection unit 1300 performs a role of selecting an optimal model capable of recommending a smart contract tag by comparing the data learned from the smart contract recommendation model learning unit with the evaluation data set.

In addition, since it is necessary to propose a plurality of tags by analyzing a pattern between one smart contract information and a plurality of tag information, the tag recommended as a result through the system of the present invention, a recommendation system algorithm is used. It is desirable.

Specifically, the result of the data learned from the plurality of models and the evaluation data set are compared to select a model with the best performance among the plurality of models. In the case of the selected model, it is applied to the smart contract tag recommendation unit 2000 to be described later.

That is, as described above, when a model having the best performance is selected by the recommendation system algorithm or the like, the model can be called as an optimal model. Although not described in detail, other types of performance evaluation models may be introduced and applied instead of the recommended system algorithm.

In addition, the smart contract tag recommendation unit 2000 shown in the lower part of FIG. 2 may include a second smart contract analysis unit 2100, a smart contract model application unit 2200, and a tag recommendation application unit 2300.

Specifically, the second smart contract analysis unit 2100 serves to extract a plurality of parameters from a smart contract requested by a user.

The smart contract discussed here has a different meaning from the smart contract mentioned in the above-described first smart contract analysis unit 1100, and is a smart contract that the user wants to request for analysis, and in reality, it is mostly in the form of binary code without source code. will be.

Therefore, the parameter extracted by the second smart contract analysis unit 2100 will be in the form of an executable code extracted from a binary code, and the resulting extracted information is preferably a component, context information, and length information.

After that, the extracted information is concatenated and then vectorized.

In addition, the smart contract model application unit 2200 applies the optimal model selected from the smart contract optimal model selection unit 1300 to the parameters (extracted information) extracted from the second smart contract analysis unit 2100. Play a role.

In addition, the tag recommendation application unit 2300 serves to recommend a tag selected from the smart contract model application unit 2200.

Specifically, the tag recommendation application unit 2300 applies a Top-k selection technique that extracts only the top k tags, and as a result, a tag to be recommended is selected and matched to a user.

Finally, for Deep Neural Networks, Deep Collaborative Filtering, and K-Nearest Neighbors (KNN) presented as an embodiment of the specific learning model described above, while explaining deep collaboration filtering in FIG. , Please explain in detail.

In Figure 2, the operation section of each of these components is shown by dividing it with a dotted line, and for a part where there is a difference in the representation of detailed functions from the drawing, first refer to the specific content for carrying out the invention. To

3 is a conceptual diagram of a deep collaborative filtering model.

FIG. 3 illustrates deep collaborative filtering used in the present invention among the plurality of models described above. Deep collaborative filtering is a method in which a deep learning algorithm is applied to collaborative filtering, which is characterized by recommending using user evaluation information. The characteristic is that the additional information representing the user and the additional information on the item are considered together with rating information for each item selected by the user. In addition, in the present invention, a marginalized denoising autoencoder, which is a method of minimizing data noise among deep learning algorithms and re-expressing them as meaningful low-dimensional vectors, is used to determine execution code information and tag information. It can be used for analysis by re-expressing it as an arbitrary f-dimensional vector that minimizes. By using such an auto encoder, the accuracy is higher and more precise results are produced than when not used. As a result, a reduced f-dimensional vector is calculated by applying an AutoEncoder to each of the execution code information and the tag information, and the product of the two vectors is calculated to indicate how appropriate the tag information is to the execution code information. Can be calculated.

Also, I would like to amplify the above-described ambient noise removal auto encoder. Basically, it is a deep learning model composed of an autoencoder, one feedforward layer (encoder), and the reversed layer (decoder). In other words, by setting the input value and the output value to the same data, the data is encoded in a dimension less than the dimension of the input value, and the original data can be calculated by re-decoding from the encoded value. It is characterized by design. In this process, the data converted to a smaller dimension can be seen as an abbreviated form while containing information about the input value. The ambient noise removal auto encoder is a modified algorithm of the auto encoder and basically follows the network structure of the auto encoder, but it inputs data by adding random noise to the original data as an input value, and calculates the original data as an output value. It is an algorithm that learns to minimize noise by itself.

Next, deep neural networks proposed as another embodiment of a specific learning model will be described.

More specifically, in the present invention, it is preferable to use a deep neural network method to which a ranking loss function is further applied.

It is a model that uses a ranking loss function as a loss function in a network composed of a feed forward layer among deep learning layers. Here, the ranking loss function is calculated differently from the general loss function and solves a problem that may occur when the parameters of the model are updated. Specifically, in the present invention, when recommending a tag, a top-k selection process is performed. That is, if only the top k tags are selected and the loss is calculated, a problem in which it is difficult to update the parameter can be solved because differential calculation is not performed in the process of selecting top-k. In addition, the trained model receives execution code information as input and passes through a deep learning algorithm (matrix operation) to calculate a vector equal to the total number of tags, which soon plays the same role as a rating.

Finally, a description will be given of KNN (K-Nearest Neighbors) proposed as another embodiment of a specific learning model.

KNN measures the similarity between each smart contract, selects k similar smart contracts, and provides tags that are frequently used within the selected group in order. In the case of such KNN modeling, there is an advantage that the learning process is very short since there is no additional model parameter. On the other hand, KNN has a disadvantage of showing relatively low performance compared to other modeling techniques in terms of speed because it requires repetitive similarity measurement calculations in the inference process.

In addition, the user device 700 mentioned in the present invention includes both a device for which a user requests a smart contract to the system of the present invention and a device for accepting a smart contract from a blockchain system. That is, the user device 700 may be a smart phone, a pad, a notebook capable of accessing the Internet, and so on, and in a larger view, all computing devices in a desktop environment may be included.

Finally, an embodiment in which the present invention is applied to the analysis of an actual smart contract will be described.

In the actual application process of the present invention, among the smart contracts stored in the Ethereum blockchain, 34,397 smart contracts containing title and source code information were extracted and used as example data. At this time, 34,397 smart contract titles and words included in function names extracted and analyzed are about 16,000, of which words that can represent the contract contents are selected through processing such as correcting typos and merging similar words through a total investigation. As a result, it was confirmed that the tag was composed of 184 words. At this time, the criteria for selecting words were selected as words representing the purpose of the smart contract, such as ICO and SALE, and words that can describe specific actions such as Refund and Freeze. Thereafter, in the process of training a plurality of models of the present invention, only 27,517 data sets corresponding to 80% of the total 34,397 data sets are used to learn, and 6,880 data sets corresponding to the remaining 20% are used in the performance evaluation process. I did. By separating the data in this way, the system evaluation was made more fair, and it was possible to evaluate the robustness of the constructed learning modeling of the present invention.

At this time, for the performance evaluation of the trained model, 10 tags are generated for each smart contract by using pre-categorized test smart contract data, and the performance of the model is compared using the performance evaluation index of the following recommendation system. Therefore, the algorithm showing the best performance was used to recommend appropriate tags in the user's smart contract.

Individual or selectively combined components constituting the automated smart contract recommendation system according to an embodiment of the present invention may be implemented as a computer program stored in a storage medium for execution through a combination with a computer.

In addition, individual or selectively combined components constituting the automated smart contract recommendation system according to an embodiment of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. have. The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes magneto-optical media, and hardware devices specially configured to store and execute program instructions such as ROM, RAM, flash memory, solid state drive (SSD), and the like. Examples of the program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

In addition, a configuration such as a computer or a computer program used in the present invention has the same meaning as a smart phone or an application running on a smart phone, as the shape of the mobile communication terminal is transformed like a smart phone, and the computing power is dramatically increased. Can be used.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art will variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the following claims. You will understand that you can do it.

10000: Smart contract recommendation system
1000: Smart contract tag recommendation model build-up unit
1100: first smart contract analysis unit
1200: Smart contract recommendation model learning unit
1300: Smart contract optimal model selection unit
2000: Smart contract tag recommendation department
2100: second smart contract analysis unit
2200: Smart contract model application unit
2300: tag recommendation application unit
700: user device

Claims (5)

In the smart contract recommendation system 10000 that recommends tags by analyzing a user's smart contract,
The recommendation system 10000
A first smart contract analysis unit 1100 for generating a plurality of learning and evaluation data sets by extracting a plurality of parameters from a plurality of smart contracts to be learned;
A smart contract recommendation model learning unit 1200 for receiving the training data set from a first smart contract analysis unit and learning it using a specific learning model;
And a smart contract optimal model selection unit 1300 that compares data learned from the smart contract recommendation model learning unit with the evaluation data set, and selects a model capable of recommending a smart contract tag,
The plurality of parameters extracted from the smart contract use the executable code (opcode) disassembled the binary code of the smart contract, and information extracted from the source code of the smart contract,
The information extracted from the execution code (opcode) includes a length, a context, and a component, and the information extracted from the source code includes a title (contract name) or a function name (function name) of the smart contract,
The component of the execution code is a smart contract recommendation system, characterized in that implemented by a one-hot encoding (one-hot encoding) technique.
In the smart contract recommendation system 10000 that recommends tags by analyzing a user's smart contract,
The recommendation system 10000
A smart contract optimal model selection unit 1300 for extracting a plurality of parameters from a plurality of smart contracts and recommending a tag through learning a plurality of models;
A second smart contract analysis unit 2100 for a user to extract a plurality of parameters from the smart contract;
A smart contract model application unit 2200 for applying the model selected from the smart contract optimal model selection unit 1300 to the parameters extracted from the second smart contract analysis unit 2100;
And a tag recommendation application unit 2300 for recommending a tag selected from the smart contract model application unit 2200,
The plurality of parameters extracted from the smart contract use the executable code (opcode) disassembled the binary code of the smart contract, and information extracted from the source code of the smart contract,
The information extracted from the execution code (opcode) includes a length, a context, and a component, and the information extracted from the source code includes a title (contract name) or a function name (function name) of the smart contract,
The component of the execution code is a smart contract recommendation system, characterized in that implemented by a one-hot encoding (one-hot encoding) technique.
delete The method of claim 1,
The specific learning model consists of a plurality of models, and each of the plurality of models selectively combines at least two of Deep Neural Networks, Deep Collaborative Filtering, and K-Nearest Neighbors (KNN). Smart contract recommendation system, characterized in that to use.
delete

Images