KR20240030078A - Blockchain smart contract vulnerability location detection method and device - Google Patents

Abstract

A method and device for detecting the location of blockchain smart contract vulnerabilities are disclosed. According to an embodiment of the present invention, a method for detecting the location of a vulnerability in a blockchain smart contract includes converting a smart contract into bytecode and performing data preprocessing; Vulnerability detection by inputting the byte code into a deep learning model; And when 'vulnerability occurrence' is detected as an output from the deep learning model, it may include a code mapping step to specify the location of the vulnerability within the smart contract.

Description

Blockchain smart contract vulnerability location detection method and device {BLOCKCHAIN SMART CONTRACT VULNERABILITY LOCATION DETECTION METHOD AND DEVICE}

The present invention uses Grad-CAM among representative XAI techniques to provide an explanation of the results of deep learning without additional data or learning, and uses image conversion and deep learning through preprocessing techniques to increase the detection rate of XAI-based smart contract vulnerability locations. This is about a blockchain smart contract vulnerability location detection method and device that provides a learning model structure.

In addition, the present invention provides a blockchain smart contract vulnerability location detection method that performs data preprocessing, vulnerability detection, and code mapping to accurately identify the location of vulnerabilities in the smart contract, and Provides a device.

Ethereum's smart contract, as a programming language, has expanded the use of blockchain to various industrial fields such as distribution, insurance, medicine, energy, and games.

However, smart contracts have various vulnerabilities that do not occur in conventional programming languages.

The reentrancy vulnerability, a representative vulnerability, took advantage of a vulnerability in smart contracts and caused a loss of more than 100 billion won to smart contract users.

Reentracy can be a hacking method that confuses smart contracts by executing multiple transactions at the same time, and then extracts funds through other transactions before certain transactions are sequentially executed.

Smart contract vulnerabilities such as reentrancy are directly linked to the leakage of assets such as cryptocurrency, so developers must be able to detect and correct these vulnerabilities in advance.

Various vulnerability analysis tools have been proposed to detect vulnerabilities, but static analysis and dynamic analysis tools, which require vulnerability detection rules and patterns to be specified in advance, can be used even if they deviate slightly from the predefined patterns. There was a problem that vulnerabilities were not detected.

To solve this problem, machine learning-based smart contract detection technologies have recently been proposed.

Deep learning, a representative machine learning technique, shows strong detection performance in vulnerability detection, but the hundreds of complex layers and millions of neurons that make up deep learning have the disadvantage of making it difficult to explain the detection results of deep learning. exist.

Due to these characteristics of deep learning, a general deep learning-based smart contract detection model can only provide results regarding vulnerabilities, and it is difficult to explain the cause of the detection results (vulnerability location, etc.).

In vulnerability detection, the difficulty in explaining the results causes usability problems that do not inform developers of the location of vulnerabilities, and reliability problems with the detection results.

Therefore, there is an urgent need for an improved detection model for vulnerabilities in smart contracts that can detect vulnerabilities and explain the causes of detection results.

An embodiment of the present invention aims to provide a blockchain smart contract vulnerability location detection method and device that detects the vulnerability location through eXplainable AI (XAI).

Additionally, the embodiment of the present invention aims to provide an explanation of the results of deep learning without additional data or learning, using Grad-CAM among the XAI techniques.

Additionally, the embodiment of the present invention aims to increase the XAI-based smart contract vulnerability location detection rate by proposing image conversion and deep learning model structure through preprocessing techniques.

According to an embodiment of the present invention, a method for detecting the location of a vulnerability in a blockchain smart contract includes converting a smart contract into bytecode and performing data preprocessing; Vulnerability detection by inputting the byte code into a deep learning model; And when 'vulnerability occurrence' is detected as an output from the deep learning model, it may include a code mapping step to specify the location of the vulnerability within the smart contract.

In addition, a blockchain smart contract vulnerability location detection device according to an embodiment of the present invention includes a data preprocessing unit that converts a smart contract into bytecode and performs data preprocessing; A vulnerability detection unit that detects vulnerabilities by inputting the byte code into a deep learning model; And when 'occurrence of vulnerability' is detected as an output from the deep learning model, it can be configured to include a code mapping unit that performs code mapping to specify the location of the vulnerability in the smart contract.

According to an embodiment of the present invention, it is possible to provide a blockchain smart contract vulnerability location detection method and device that detects the vulnerability location through eXplainable AI (XAI).

In addition, according to the present invention, an explanation of the results of deep learning can be provided without additional data or learning by using Grad-CAM among the XAI techniques.

In addition, the present invention can increase the XAI-based smart contract vulnerability location detection rate by proposing image conversion and deep learning model structure through preprocessing techniques.

Figure 1 is a block diagram showing the configuration of a blockchain smart contract vulnerability location detection device according to an embodiment of the present invention.
Figure 2 is a diagram explaining the operation sequence of the blockchain smart contract vulnerability location detection device.
Figure 3 is a diagram for explaining byte code conversion through Transforming.
Figure 4 is a flowchart showing a method for detecting the location of a blockchain smart contract vulnerability according to an embodiment of the present invention.

Hereinafter, embodiments will be described in detail with reference to the attached drawings. However, various changes can be made to the embodiments, so the scope of the patent application is not limited or limited by these embodiments. It should be understood that all changes, equivalents, or substitutes for the embodiments are included in the scope of rights.

The terms used in the examples are for descriptive purposes only and should not be construed as limiting. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as “comprise” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but are not intended to indicate the presence of one or more other features. It should be understood that this does not exclude in advance the possibility of the existence or addition of elements, numbers, steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the technical field to which the embodiments belong. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and unless explicitly defined in the present application, should not be interpreted in an ideal or excessively formal sense. No.

In addition, when describing with reference to the accompanying drawings, identical components will be assigned the same reference numerals regardless of the reference numerals, and overlapping descriptions thereof will be omitted. In describing the embodiments, if it is determined that detailed descriptions of related known technologies may unnecessarily obscure the gist of the embodiments, the detailed descriptions are omitted.

Figure 1 is a block diagram showing the configuration of a blockchain smart contract vulnerability location detection device according to an embodiment of the present invention.

Referring to FIG. 1, the blockchain smart contract vulnerability location detection device 100 according to an embodiment of the present invention includes a data preprocessing unit 110, a vulnerability detection unit 120, and a code mapping unit 130. It can be configured to include.

First, the data preprocessing unit 110 converts the smart contract into bytecode and performs data preprocessing. In other words, the data preprocessing unit 110 can serve to preprocess the smart contract in a form that allows input into a deep learning model related to detecting whether there are vulnerabilities in the smart contract.

Data pre-processing by the data pre-processing unit 110 may consist of a Compile step, a Transforming step, and a Mapping step.

In the Compile step, the data preprocessor 110 can compile the smart contract, which is a high-level language, into the byte code, which is a low-level language.

A smart contract can refer to a contract using a software program in which a transaction is automatically concluded between parties when certain conditions are met in an environment based on Distributed Ledger Technology (DLT).

Figure 2, described later, illustrates a smart contract (contract SimpleDAO {~}) described in the high-level language of a software program.

Bytecode can be a binary representation for an executable program that runs on a virtual computer rather than on specific hardware.

Figure 2 illustrates that the above-described smart contract (contract SimpleDAO {~}) is compiled into bytecodes (Bytecodes: 0x608062~).

In the Transforming step, the data preprocessor 110 may transform the compiled byte code into an input form specified in the deep learning model.

In other words, the data preprocessor 110 serves to unify the structure of the byte code into a form that can be input into a deep learning model.

Transformation can be accomplished by removing operands and adding zero padding.

The data preprocessor 110 can check the number of operands assigned to each opcode in the byte code.

For example, in an image-based deep learning model (CNN), two operands may be allocated to each opcode for a byte code of the size required.

If the byte code Original Opcode consists of '60(Opcode)-80(Operand)/62(Opcode)-40,50,78(Operand)/52(Opcode)', the data preprocessor 110 You can check '60(Opcode)-1 Operand allocated'/62(Opcode)-3 Operand allocated)/52(Opcode)-No Operand allocated'.

If the number of confirmed operands is greater than the number specified in the deep learning model, the data preprocessor 110 may remove the exceeding number of operands.

For '62(Opcode)-40, 50, 78(Operand)' to which three operands are allocated, the data preprocessor 110 can remove the last operand 78 that exceeds the specified two. Accordingly, '62(Opcode)' can be assigned two Operands 40 and 50 according to regulations.

If the number of confirmed operands is less than the number specified in the deep learning model, zero padding equal to the insufficient number can be added.

For '60(Opcode)-80(Operand)' assigned to one Operand, the data pre-processing unit 110 may add zero padding 00 for one insufficient number to match the specified two. Accordingly, '60(Opcode)' can be assigned two Operands 80,00 in accordance with regulations.

In addition, for '52 (Opcode)' for which no operand is assigned, the data preprocessor 110 may add two insufficient zero paddings, 00 00, to match the specified two. Accordingly, '52(Opcode)' can be assigned two Operands 00,00 in accordance with regulations.

Finally, the data preprocessor 110 sets the byte to '60(Opcode)-80,00(Operand)/62(Opcode)-40,50(Operand)/52(Opcode)-00,00(Operand)'. The original opcode can be converted into a form suitable for an image-based deep learning model (CNN).

In the mapping step, the data preprocessor 110 may map the converted byte code to an RGB image. That is, the data preprocessor 110 can map the byte code and the RGB image 1:1, allowing the byte code to be input to an image-based deep learning model (CNN) as a deep learning model and learn.

The vulnerability detection unit 120 inputs the byte code into a deep learning model to detect vulnerabilities. In other words, the vulnerability detection unit 120 can ultimately serve to detect whether there is a vulnerability in the smart contract based on learning from a deep learning model for the imaged byte code.

The vulnerability detection unit 120 can detect the presence or absence of vulnerabilities by using the RGB image obtained from data preprocessing as an input to a deep learning model.

To detect the presence or absence of vulnerabilities, the deep learning model according to the present invention may include a minimized pooling and non-stride structure, a Depthwise Separable Convolution structure, and a Global Max Pooling layer.

That is, the vulnerability detection unit 120 may not use the pooling operation and stride operation in the deep learning model.

Pooling and stride structures can be used in general deep learning models to reduce computation and increase learning efficiency.

Pooling, for example, can be a technique to reduce the amount of computation by representing multiple pixels as one pixel (the pixel with the most influence or the pixel with the least influence).

The stride structure can be a technique to reduce the size of data by arbitrarily reducing a plurality of data by one piece of data.

However, pooling and stride operations may cause the problem of losing vulnerability spatial information during the learning and detection process.

Accordingly, the vulnerability detection unit 120 can prevent pooling operations and stride operations from being used by using a deep learning model with a minimal pooling and non-stride structure in order to preserve vulnerability location information.

The vulnerability detection unit 120 can reduce the number of parameters consumed in the convolution operation by using the Depthwise Separable Convolution operation in the deep learning model instead of the pooling operation and the stride operation.

Due to the use of deep learning models that do not use pooling or stride operations, the learning process may require a large amount of computation.

To solve this, the vulnerability detection unit 120 can use a Depthwise Separable Convolution operation that can reduce the number of parameters consumed in the convolution operation by using a deep learning model with a Depthwise Separable Convolution structure.

Depthwise Separable Convolution operation for RGB images can reduce memory usage and increase learning speed by optimizing the number of parameters. In the Depthwise Separable Convolution operation, for example, the Depthwise Separable Convolution operation can be performed with two types of filters separated into ‘filter height, filter width, 1’ and ‘1, 1, number of channels’, respectively, based on the channel.

Depthwise Separable Convolution operation may be a method of decomposing depth and performing multiple convolution operations.

In addition, the vulnerability detection unit 120 can compensate for the decrease in computational efficiency caused by the addition of the zero padding by using the Global Max Pooling layer as the last layer of the deep learning model.

In general, Fully Connected Layer, Global Average Pooling, etc. can be used to detect vulnerabilities in the last layer of deep learning.

However, since zero padding was added in the transforming stage of data preprocessing, the efficiency of calculation may be reduced when using Fully Connected Layer or Global Average Pooling.

To solve the problem of reduced computational efficiency, the vulnerability detection unit 120 can use the last layer of the deep learning model as a Global Max Pooling layer.

Pooling is mainly performed after the convolution operation, and Global Max Pooling can be a layer for extracting the maximum value from the window of the result of the Depthwise Separable Convolution operation.

In addition, when 'occurrence of vulnerability' is detected as an output from the deep learning model, the code mapping unit 130 performs code mapping to specify the location of the vulnerability within the smart contract. That is, the code mapping unit 130 can serve to identify the location of the detected vulnerability and then restore it to a smart contract, a high-level language, through code mapping to explain the location of the vulnerability.

In code mapping, the code mapping unit 130 can find the location of vulnerabilities in the RGB image through the application of Grad-CAM. In other words, the code mapping unit 130 can find vulnerabilities in the byte code used as input through the application of Grad-CAM.

Grad-CAM can be an activation mapping technique based on gradient.

The code mapping unit 130 can use Grad-CAM to classify RGB images into classes and localize and express predicted vulnerabilities in the RGB images. In other words, the code mapping unit 130 can accurately find the exact location of the vulnerability by localizing and expressing the location of the vulnerability through Grad-CAM.

Additionally, the code mapping unit 130 may map the found vulnerability location to an instruction of the byte code through byte code mapping. That is, the code mapping unit 130 can convert the vulnerability location found in the RGB image into a byte code language.

Thereafter, the code mapping unit 130 may map the commands of the byte code to the commands of the smart contract through smart contract code mapping. In other words, the code mapping unit 130 can convert the byte code, which is a low-level language, into a smart contract, which is a high-level language, so that the location of the discovered vulnerability is finally expressed in a language that the user can recognize.

According to an embodiment of the present invention, it is possible to provide a blockchain smart contract vulnerability location detection method and device that detects the vulnerability location through eXplainable AI (XAI).

In addition, according to the present invention, an explanation of the results of deep learning can be provided without additional data or learning by using Grad-CAM among the XAI techniques.

In addition, the present invention can increase the XAI-based smart contract vulnerability location detection rate by proposing image conversion and deep learning model structure through preprocessing techniques.

Existing deep learning-based smart contract detection techniques have the disadvantage of being difficult to detect vulnerability locations.

There have been some attempts to provide detection of vulnerability locations, but these studies have the disadvantage of requiring additional structures such as vulnerability location data or layers in the learning stage.

To solve this problem, the blockchain smart contract vulnerability location detection device 100 of the present invention proposes a method of detecting vulnerability locations through eXplainable AI (XAI).

The blockchain smart contract vulnerability location detection device 100 uses Grad-CAM, one of the representative XAI techniques, to provide an explanation of the results of deep learning without additional data or learning.

In addition, the blockchain smart contract vulnerability location detection device 100 proposes image conversion and a deep learning model structure through preprocessing techniques to increase the XAI-based smart contract vulnerability location detection rate.

Figure 2 is a diagram explaining the operation sequence of the blockchain smart contract vulnerability location detection device.

The blockchain smart contract vulnerability location detection device 100 can sequentially perform data preprocessing (STEP1. Data Preprocessing), vulnerability detection (STEP2. Vulnerability Detection), and code mapping (STEP3. Code Mapping).

The blockchain smart contract vulnerability location detection device 100 can receive smart contract source code as input and perform data preprocessing to convert it into an input form suitable for a deep learning model.

The blockchain smart contract vulnerability location detection device 100 can detect vulnerabilities in a deep learning model using data preprocessed smart contract source code.

The blockchain smart contract vulnerability location detection device 100 can detect the vulnerability location by performing code mapping to find the vulnerability location when a vulnerability occurs.

As shown in Figure 2, data preprocessing may consist of a Compile step, a Transforming step, and a Mapping step.

Data preprocessing converts the Solidity smart contract, a high-level language, into bytecode, a low-level language, through compilation, and then uses it for learning in a deep learning model. It can be converted into a suitable form.

Afterwards, the blockchain smart contract vulnerability location detection device 100 can complete data preprocessing by converting it into an image-based contract to use it as input to a deep learning model.

The Compile stage of data preprocessing may be the process of converting a smart contract into byte code, a low-level language.

The transforming step of data preprocessing may be a process of converting byte code into a form suitable for learning. An image-based deep learning model (CNN) can use bytecodes of a fixed size for calculation, for example, as shown in Figure 3 described later, a size suitable for deep learning learning, and requires up to two arguments (Operand) per bytecode operation. It can be converted to receive only .

For bytecode operations that take no arguments, zero padding can be added.

Figure 3 is a diagram for explaining byte code conversion through Transforming.

In Figure 3, a byte code of the size required for an image-based deep learning model (CNN) is illustrated, consisting of an Opcode and two Operands.

As shown in Figure 3, when the Original Opcode consists of '60(Opcode)-80(Operand)/62(Opcode)-40,50,78(Operand)/52(Opcode)', there is a blockchain smart contract vulnerability. The location detection device 100 can remove more than two Operands 78 from '62(Opcode)-40,50,78(Operand)'.

Additionally, the blockchain smart contract vulnerability location detection device 100 can add zero padding to each of '60(Opcode)-80(Operand)' and '52(Opcode)' as required.

Finally, the blockchain smart contract vulnerability location detection device 100 is composed of an opcode and two operands by removing the operand and adding zero padding, '60(Opcode)-80,00(Operand)/62(Opcode )-40,50(Operand)/52(Opcode)-00,0(Operand)' can be converted to a byte code of the size required by the image-based deep learning model (CNN).

The mapping step of data preprocessing may be a process of imaging the converted byte code.

The byte code can have one of the numbers from 0 to 255, and the RGB image can also have one number from 0 to 255, so the blockchain smart contract vulnerability location detection device 100 uses the byte code and RGB Images can be mapped 1:1.

Vulnerability detection may be the process of detecting the occurrence of a vulnerability through a deep learning model with a minimized pooling and non-stride structure, a Depthwise Separable Convolution structure, and a Global Max Pooling layer.

In detecting vulnerabilities, the blockchain smart contract vulnerability location detection device 100 can detect the presence or absence of vulnerabilities by using the RGB image obtained from data preprocessing as an input to a deep learning model.

If a vulnerability is discovered, the blockchain smart contract vulnerability location detection device 100 can perform subsequent vulnerability location detection for the discovered vulnerability.

If no vulnerability is found, the blockchain smart contract vulnerability location detection device 100 may terminate processing.

Deep learning models have minimized pooling and non-stride structures.

In general deep learning models, pooling or stride operations can be used to reduce the amount of computation and increase learning efficiency.

However, these operations have the problem of losing vulnerability spatial information during the learning and detection process.

Therefore, the deep learning model used in the present invention can use minimal pooling and non-stride structures to preserve vulnerability location information.

Deep learning models have a Depthwise Separable Convolution structure.

Because the deep learning model used in the present invention does not use pooling or stride operations, a large amount of calculations are required during the deep learning process.

To solve this problem, deep learning models can use Depthwise Separable Convolution operation, which can reduce the number of parameters consumed in convolution operation.

The deep learning model has a Global Max Pooling layer.

In general, Fully Connected Layer, Global Average Pooling, etc. can be used to detect vulnerabilities in the last layer of deep learning.

However, since zero padding was added in the transforming stage of data preprocessing, the efficiency of calculation may be reduced when using Fully Connected Layer or Global Average Pooling.

To solve the problem of reduced computational efficiency, deep learning models can use the Global Max Pooling layer as the last layer.

Code mapping can perform Grad-CAM-based vulnerability location detection, byte code mapping, and smart contract code mapping (Solidity).

Code mapping can be a process of finding vulnerability locations in an RGB image through Grad-CAM, an XAI technology, and converting and expressing the found vulnerability locations into byte code, a low-level language, and smart contract code, a high-level language.

Code Mapping's Grad-CAM-based vulnerability location detection is the process of finding where vulnerabilities are in the RGB image used as input through the application of Grad-CAM.

Code Mapping's byte code mapping is the process of mapping the discovered vulnerability location to instructions on the byte code.

Code Mapping's smart contract code mapping (Solidity) is the process of mapping the commands of the mapped byte code into the Solidity language, an existing high-level language.

According to the present invention, unlike existing simple deep learning (CNN)-based smart contract detection techniques, it is possible to detect vulnerability locations through a deep learning model.

According to the present invention, vulnerability locations can be detected without learning additional data or consuming layers.

According to the present invention, an explanation of the results of a deep learning model can be provided through code mapping, and information for improving the performance of the deep learning model can be provided.

Hereinafter, in FIG. 4, the work flow of the blockchain smart contract vulnerability location detection device 100 according to embodiments of the present invention will be described in detail.

Figure 4 is a flowchart showing a method for detecting the location of a blockchain smart contract vulnerability according to an embodiment of the present invention.

The blockchain smart contract vulnerability location detection method according to this embodiment can be performed by the blockchain smart contract vulnerability location detection device 100.

First, the blockchain smart contract vulnerability location detection device 100 converts the smart contract into bytecode and performs data preprocessing (410). Step 410 may be a process of preprocessing the smart contract in a form that allows input to a deep learning model related to detecting whether there is a vulnerability in the smart contract.

Data preprocessing by the blockchain smart contract vulnerability location detection device 100 may consist of a Compile step, a Transforming step, and a Mapping step.

In the Compile step, the blockchain smart contract vulnerability location detection device 100 compiles the smart contract, which is a high-level language, into the byte code, which is a low-level language. can do.

A smart contract can refer to a contract using a software program that automatically concludes a transaction between parties when certain conditions are met in an environment based on Distributed Ledger Technology (DLT).

Figure 2 illustrates a smart contract (contract SimpleDAO {~}) described in the high-level language of a software program.

Bytecode can be a binary representation for an executable program that runs on a virtual computer rather than on specific hardware.

Figure 2 illustrates that the above-described smart contract (contract SimpleDAO {~}) is compiled into bytecodes (Bytecodes: 0x608062~).

In the Transforming step, the blockchain smart contract vulnerability location detection device 100 can transform the compiled byte code into an input form specified in the deep learning model.

In other words, the blockchain smart contract vulnerability location detection device 100 serves to unify the structure of the byte code into a form that can be input into a deep learning model.

Transformation can be accomplished by removing operands and adding zero padding.

The blockchain smart contract vulnerability location detection device 100 can check the number of Operands assigned to each Opcode in the byte code.

For example, in an image-based deep learning model (CNN), two operands may be allocated to each opcode for a byte code of the size required.

If the byte code Original Opcode is composed of '60(Opcode)-80(Operand)/62(Opcode)-40,50,78(Operand)/52(Opcode)', the blockchain smart contract vulnerability location detection device ( 100), you can check '60(Opcode)-1 Operand allocated'/62(Opcode)-3 Operand allocated)/52(Opcode)-No Operand allocated'.

If the number of confirmed operands is greater than the number specified in the deep learning model, the blockchain smart contract vulnerability location detection device 100 can remove the exceeding number of operands.

For '62(Opcode)-40, 50, 78(Operand)' assigned to 3 operands, the blockchain smart contract vulnerability location detection device 100 can remove the last operand 78 that exceeds the specified 2. there is. Accordingly, '62(Opcode)' can be assigned two Operands 40 and 50 according to regulations.

If the number of confirmed operands is less than the number specified in the deep learning model, zero padding equal to the insufficient number can be added.

For '60(Opcode)-80(Operand)' assigned to one Operand, the blockchain smart contract vulnerability location detection device (100) can add zero padding 00 for one insufficient number to match the specified two. there is. Accordingly, '60(Opcode)' can be assigned two Operands 80,00 in accordance with regulations.

In addition, for '52 (Opcode)' for which no operand is assigned, the blockchain smart contract vulnerability location detection device 100 can add the insufficient two zero paddings 00 00 to match the specified two. Accordingly, '52(Opcode)' can be assigned two Operands 00,00 in accordance with regulations.

Finally, the blockchain smart contract vulnerability location detection device 100 is '60(Opcode)-80,00(Operand)/62(Opcode)-40,50(Operand)/52(Opcode)-00,00(Operand )', the byte code Original Opcode can be converted into a form suitable for an image-based deep learning model (CNN).

In the Mapping step, the blockchain smart contract vulnerability location detection device 100 can map the converted byte code to an RGB image. In other words, the blockchain smart contract vulnerability location detection device 100 maps the byte code and RGB image 1:1, allowing the byte code to be input into an image-based deep learning model (CNN) as a deep learning model and learn. there is.

In addition, the blockchain smart contract vulnerability location detection device 100 inputs the byte code into a deep learning model to detect vulnerabilities (420). Step 420 may be a process of ultimately detecting whether there is a vulnerability in the smart contract based on learning in a deep learning model for the imaged byte code.

The blockchain smart contract vulnerability location detection device 100 can detect the presence or absence of vulnerabilities by using the RGB image obtained from data preprocessing as an input to a deep learning model.

To detect the presence or absence of vulnerabilities, the deep learning model according to the present invention may include a minimized pooling and non-stride structure, a Depthwise Separable Convolution structure, and a Global Max Pooling layer.

In other words, the blockchain smart contract vulnerability location detection device 100 may not use the pooling operation and stride operation in the deep learning model.

Pooling and stride structures can be used in general deep learning models to reduce computation and increase learning efficiency.

However, pooling and stride operations may cause the problem of losing vulnerability spatial information during the learning and detection process.

Accordingly, the blockchain smart contract vulnerability location detection device 100 can prevent pooling operations and stride operations from being used by a deep learning model with a minimal pooling and non-stride structure in order to preserve vulnerability location information. there is.

The blockchain smart contract vulnerability location detection device 100 can reduce the number of parameters consumed in the convolution operation by using the Depthwise Separable Convolution operation in the deep learning model instead of the pooling operation and the stride operation.

Due to the use of deep learning models that do not use pooling or stride operations, the learning process may require a large amount of computation.

To solve this, the blockchain smart contract vulnerability location detection device 100 can use a Depthwise Separable Convolution operation that can reduce the number of parameters consumed in the convolution operation by using a deep learning model with a Depthwise Separable Convolution structure.

Depthwise Separable Convolution operation for RGB images can reduce memory usage and increase learning speed by optimizing the number of parameters. In the Depthwise Separable Convolution operation, for example, the Depthwise Separable Convolution operation can be performed with two types of filters separated into ‘filter height, filter width, 1’ and ‘1, 1, number of channels’, respectively, based on the channel.

In addition, the blockchain smart contract vulnerability location detection device 100 can compensate for the decrease in computational efficiency caused by the addition of the zero padding by using the Global Max Pooling layer as the last layer of the deep learning model.

In general, Fully Connected Layer, Global Average Pooling, etc. can be used to detect vulnerabilities in the last layer of deep learning.

However, since zero padding was added in the transforming stage of data preprocessing, the efficiency of calculation may be reduced when using Fully Connected Layer or Global Average Pooling.

To solve the problem of reduced computational efficiency, the blockchain smart contract vulnerability location detection device 100 can use the last layer of the deep learning model as the Global Max Pooling layer.

Pooling is mainly performed after the convolution operation, and Global Max Pooling can be a layer for extracting the maximum value from the window of the result of the Depthwise Separable Convolution operation.

Continuing, when the blockchain smart contract vulnerability location detection device 100 detects 'vulnerability occurrence' as an output from the deep learning model, code mapping is performed to specify the location of the vulnerability within the smart contract (430) ). Step 430 may be a process of identifying the location of the detected vulnerability and then restoring it to a smart contract, a high-level language, through code mapping to explain the location of the vulnerability.

In code mapping, the blockchain smart contract vulnerability location detection device 100 can find the vulnerability location in the RGB image through the application of Grad-CAM. In other words, the blockchain smart contract vulnerability location detection device 100 can find where the vulnerability is in the byte code used as input through the application of Grad-CAM.

Grad-CAM can be an activation mapping technique based on gradient.

The blockchain smart contract vulnerability location detection device 100 can use Grad-CAM to distinguish classes for RGB images and localize and express the predicted vulnerability location in the RGB image. In other words, the blockchain smart contract vulnerability location detection device 100 can accurately find the exact location of the vulnerability by localizing and expressing the vulnerability location through Grad-CAM.

Additionally, the blockchain smart contract vulnerability location detection device 100 can map the found vulnerability location to an instruction of the byte code through byte code mapping. In other words, the blockchain smart contract vulnerability location detection device 100 can convert the vulnerability location found in the RGB image into a byte code language.

Afterwards, the blockchain smart contract vulnerability location detection device 100 can map the commands of the byte code to the commands of the smart contract through smart contract code mapping. In other words, the blockchain smart contract vulnerability location detection device 100 converts the byte code, which is a low-level language, into a smart contract, which is a high-level language, so that the discovered vulnerability location is finally expressed in a language that the user can recognize. there is.

According to an embodiment of the present invention, it is possible to provide a blockchain smart contract vulnerability location detection method and device that detects the vulnerability location through eXplainable AI (XAI).

In addition, according to the present invention, an explanation of the results of deep learning can be provided without additional data or learning by using Grad-CAM among the XAI techniques.

In addition, the present invention can increase the XAI-based smart contract vulnerability location detection rate by proposing image conversion and deep learning model structure through preprocessing techniques.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the medium may be specially designed and configured for the embodiment or may be known and available to those skilled in the art of 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 optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. Software and/or data may be used on any type of machine, component, physical device, virtual equipment, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

Although the embodiments have been described with limited drawings as described above, those skilled in the art can apply various technical modifications and variations based on the above. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the following claims.

100: Blockchain smart contract vulnerability location detection device
110: data preprocessing unit
120: Vulnerability detection unit
130: Code mapping unit

Claims (11)

Data preprocessing by converting the smart contract into bytecode;
Vulnerability detection by inputting the byte code into a deep learning model; and
When 'vulnerability occurrence' is detected as an output from the deep learning model, code mapping is performed to specify the location of the vulnerability within the smart contract.
Method for detecting location of blockchain smart contract vulnerabilities, including. According to paragraph 1,
The data preprocessing step is,
Compiling the smart contract, which is a high-level language, into the byte code, which is a low-level language;
Transforming the compiled byte code into an input form defined by the deep learning model; and
Mapping the converted byte code into an RGB image
Method for detecting location of blockchain smart contract vulnerabilities, including. According to paragraph 2,
The conversion step is,
Checking the number of operands assigned to each opcode in the byte code; and
If the number of confirmed operands is greater than the number specified in the deep learning model, removing the exceeding number of operands; or
If the confirmed number of Operands is less than the number specified in the deep learning model, adding zero padding equal to the insufficient number.
Method for detecting location of blockchain smart contract vulnerabilities, including. According to paragraph 3,
The step of detecting the vulnerability is,
Not using pooling and stride operations in the deep learning model;
Reducing the number of parameters consumed in a convolution operation by using a Depthwise Separable Convolution operation in the deep learning model instead of the pooling operation and the stride operation; and
Compensating for the decrease in computational efficiency due to the addition of the zero padding by using the Global Max Pooling layer as the last layer of the deep learning model.
Method for detecting location of blockchain smart contract vulnerabilities, including. According to paragraph 2,
The code mapping step is,
Finding the location of vulnerabilities in the RGB image through application of Grad-CAM;
mapping the found vulnerability location to an instruction of the byte code by byte code mapping; and
Mapping the commands of the byte code to the commands of the smart contract through smart contract code mapping.
Method for detecting location of blockchain smart contract vulnerabilities, including. A data preprocessing unit that converts smart contracts into bytecode and performs data preprocessing;
A vulnerability detection unit that detects vulnerabilities by inputting the byte code into a deep learning model; and
When a 'vulnerability occurrence' is detected as an output from the deep learning model, a code mapping unit performs code mapping to specify the location of the vulnerability within the smart contract.
Blockchain smart contract vulnerability location detection device, including. According to clause 6,
The data preprocessing unit,
Compile the smart contract, which is a high-level language, into the byte code, which is a low-level language,
Transforming the compiled byte code into an input form specified in the deep learning model,
Mapping the converted byte code into an RGB image
Blockchain smart contract vulnerability location detection device. In clause 7,
The data preprocessing unit,
Check the number of operands assigned to each opcode in the byte code,
If the number of confirmed operands is greater than the number specified in the deep learning model, the exceeding number of operands is removed,
If the number of confirmed operands is less than the number specified in the deep learning model, zero padding equal to the insufficient number is added.
Blockchain smart contract vulnerability location detection device. According to clause 8,
The vulnerability detection unit,
In the deep learning model, without using pooling operation and stride operation,
Instead of the pooling operation and the stride operation, the deep learning model uses a Depthwise Separable Convolution operation to reduce the number of parameters consumed in the convolution operation,
By using the Global Max Pooling layer as the last layer of the deep learning model, the decrease in computational efficiency caused by the addition of the zero padding is compensated.
Blockchain smart contract vulnerability location detection device. In clause 7,
The code mapping unit,
Through the application of Grad-CAM, the location of vulnerabilities is found in the RGB image,
By byte code mapping, the found vulnerability location is mapped to an instruction of the byte code,
Through smart contract code mapping, the commands of the byte code are mapped to the commands of the smart contract.
Blockchain smart contract vulnerability location detection device. A computer-readable recording medium recording a program for executing the method of any one of claims 1 to 5.

Images