Smart contract-oriented vulnerability detection system and method

Abstract

The present invention relates to a smart contract-oriented vulnerability detection system and method. The system comprises: a data preprocessing module, configured to collect a smart contract, delete useless information of the smart contract, so as to obtain a smart contract slice related to vulnerability formation, and label different vulnerability types of the smart contract slice to obtain a labeling result; a feature extraction module, configured to construct a smart contract vulnerability detection model on the basis of the labeling result, and obtain a detection result on the basis of the smart contract vulnerability detection model; and a classification module, configured to classify the detection result to complete vulnerability detection. The data preprocessing module, the feature extraction module and the classification module are connected in sequence. According to the present invention, by performing slice detection on a source code of the smart contract, the length of input information is reduced, the training workload is reduced, useful parts in a code snippet are denser, and the reliability and accuracy of the smart contract vulnerability detection system are improved.

Description

A vulnerability detection system and method for smart contracts Technical Field

[0001] The present invention relates to the field of blockchain and smart contract technologies, and in particular to a vulnerability detection system and method for smart contracts. Background Art

[0002] With the continuous development of computer technology, blockchain has gained widespread application due to its decentralized nature. Smart contracts are automated digital contracts based on blockchain technology. Deployed and run on the blockchain, they feature automatic execution and immutability, making them widely used in finance, logistics, the Internet of Things, and other fields. In recent years, national plans have also vigorously developed blockchain technology and applied it to key areas. However, due to inconsistent developer skills, many smart contracts contain vulnerabilities, resulting in significant potential economic losses and security risks for users. Therefore, vulnerability detection in smart contracts has become a critical issue in smart contract development and application. When performing vulnerability detection on smart contracts, it is important to consider the accuracy, adaptability, scalability, and reliability of the detection method. For example, in real-world applications, the code structure and logic of smart contracts may change continuously due to changing requirements. This requires rapid adaptation to these changes and timely updating of vulnerability detection models. Furthermore, the development and application of smart contracts are experiencing rapid growth, necessitating an efficient and scalable smart contract vulnerability detection system to meet the large-scale demand for smart contract detection.

[0003] Existing smart contract vulnerability detection technologies include traditional manual auditing and automated vulnerability detection. The disadvantages of manual auditing include being time-consuming, inefficient, subject to significant human interference, and inadequate for large-scale smart contract development and application. Automated smart contract vulnerability detection technologies can be further categorized into methods based on formal verification, symbolic execution, and machine learning. For example, there is a formal verification method using the F* language to verify whether smart contracts have vulnerabilities (Bhargavan K, Delignat-Lavaud A, Fournet C, et al. Formal verification of smart contracts: Short paper [C] / / Proceedings of the 2016 ACM workshop on programming languages ​​and analysis for security. 2016: 91-96.), a symbolic execution method that uses pattern matching to analyze whether smart contracts have vulnerabilities given features (Securify (Tsankov P, Dan A, Drachsler-Cohen D, et al. Securify: Practical security analysis of smart contracts [C] / / Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security. 2018: 67-82.), and a machine learning method for smart contract vulnerability detection by constructing a smart contract graph and using graph neural networks (Zhuang Y, Liu Z, Qian P, et al. Smart Contract Vulnerability Detection using Graph Neural Networks). Network[C] / / IJCAI.2020:3283-3290.). Although the above methods can improve detection efficiency to a certain extent, they still have problems such as incomplete vulnerability detection, high false positive rate, and limited scope of application.

[0004] In summary, current smart contract vulnerability detection suffers from incomplete classification and inaccurate results. This paper proposes a deep learning-based smart contract vulnerability detection model and method. By using CodeBert, BiGRU, and an attention mechanism to process the Solidity source code of smart contracts, it effectively addresses the low accuracy of vulnerability detection in smart contracts.

[0005] Summary of the Invention

[0006] The purpose of this invention is to provide a vulnerability detection system and method for smart contracts, aiming to solve the problem of low accuracy of vulnerability detection in smart contracts.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A vulnerability detection system for smart contracts, including:

[0009] A data preprocessing module is used to collect smart contracts, delete useless information from the smart contracts, obtain smart contract slices related to vulnerability formation, label the smart contract slices according to different vulnerability types, and obtain labeling results;

[0010] A feature extraction module is used to build a smart contract vulnerability detection model based on the annotation results, and obtain detection results based on the smart contract vulnerability detection model;

[0011] A classification module, used to classify the detection results and complete vulnerability detection;

[0012] The data preprocessing module, the feature extraction module and the classification module are connected in sequence.

[0013] Optionally, the data preprocessing module includes:

[0014] The contract slicing submodule is used to delete useless information from the smart contract and obtain smart contract slices related to the vulnerability;

[0015] A labeling submodule is used to label the smart contract slice with different vulnerability types and obtain the labeling results; wherein the vulnerability types include: reentrancy vulnerability, integer overflow vulnerability, denial of service vulnerability, TxOrigin vulnerability, timestamp dependency vulnerability and low-level call vulnerability;

[0016] The transmission submodule is used to transmit the annotation result to the feature extraction module.

[0017] Optionally, the feature extraction module includes:

[0018] The slice acquisition submodule is used to obtain the annotated historical smart contract slices;

[0019] The detection model construction submodule is used to form a historical word embedding matrix for the annotated historical smart contract slices through the CodeBert pre-training model, input the historical word embedding matrix into the BiGRU model of the attention mechanism for training, and obtain the smart contract vulnerability detection model.

[0020] Optionally, the detection model building submodule includes:

[0021] A pre-trained model construction unit is configured to add words that belong only to solidity to the CodeBert pre-trained model, obtain a fine-tuned CodeBert model, and input the annotated historical smart contract slices into the fine-tuned CodeBert model to form the historical word embedding matrix;

[0022] The detection model construction unit is used to input the historical word embedding matrix into the BiGRU model of the attention mechanism for training to obtain the smart contract vulnerability detection model.

[0023] Optionally, the feature extraction module further includes:

[0024] The detection result acquisition submodule is used to input the annotation results into the fine-tuned CodeBert model to form the word embedding matrix, and input the word embedding matrix into the smart contract vulnerability detection model for feature extraction to obtain a feature vector; wherein, the smart contract vulnerability detection model includes: a Bi-GRU encoding part and an attention mechanism part.

[0025] Optionally, the feature extraction module further includes:

[0026] A word embedding matrix acquisition unit, configured to input the annotation results into the fine-tuned CodeBert model to form the word embedding matrix;

[0027] A hidden layer vector acquisition unit is used to input the word embedding matrix into the Bi-GRU encoding part, obtain the output result of the forward hidden layer and the output result of the backward hidden layer, and splice the output result of the forward hidden layer and the output result of the backward hidden layer to obtain a hidden layer vector;

[0028] The feature vector acquisition unit is used to input the hidden layer vector into the attention mechanism part, configure the weights of the words in the hidden layer vector according to different importance levels, and combine them to obtain the feature vector.

[0029] Optionally, the method for obtaining the output result of the forward hidden layer and the output result of the backward hidden layer is:

[0030] in, is the output of the forward hidden layer, is the output result of the backward hidden layer, is the forward GRU unit, x t is the vector of the word, They are the forward and backward GRU unit output results at time t-1, is the backward GRU unit.

[0031] Optionally, the method for obtaining the feature vector is:

[0032] Among them, m t is the hidden layer vector mapping result, W v 、m w 、b v All are dynamic parameters. is the hidden layer vector, tanh() is the activation function, a t is the attention weight vector, softmax() is the loss function, m t T is transposed, a k for a t , L is the total length, vec is the weighted hidden state, and t is the current moment.

[0033] Optionally, the classification module includes:

[0034] The classification submodule is used to calculate the probability score of each label in the feature vector by using softmax as the activation function and the cross entropy information function as its loss function to obtain the classification result.

[0035] To achieve the above objectives, the present invention also provides a vulnerability detection method for smart contracts, comprising:

[0036] Collect smart contracts, delete useless information from the smart contracts, mark different vulnerability types, and obtain marking results;

[0037] Construct a smart contract vulnerability detection model, input the annotation results into the smart contract vulnerability detection model, obtain detection results, classify the detection results, and complete vulnerability detection.

[0038] The beneficial effects of the present invention are:

[0039] (1) By slicing the source code of the smart contract for detection, the present invention reduces the length of the input information, reduces the workload of training, makes the useful parts of the code fragments more dense, and improves the reliability and accuracy of the smart contract vulnerability detection system.

[0040] (2) The present invention identifies six different smart contract vulnerabilities by analyzing the causes of multiple vulnerabilities, and implements a multi-classification model for reentrancy vulnerabilities, integer overflow vulnerabilities, denial of service vulnerabilities, TxOrigin vulnerabilities, timestamp dependency vulnerabilities, and low-level call vulnerabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] FIG1 is a schematic diagram of a vulnerability detection system for smart contracts according to an embodiment of the present invention;

[0043] FIG2 is a flowchart of generating smart contract code slices according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] The present invention discloses a vulnerability detection system for smart contracts, comprising: a data preprocessing module for collecting smart contracts, deleting useless information from the smart contracts, labeling them according to different vulnerability types, and obtaining labeling results; a feature extraction module for building a smart contract vulnerability detection model, inputting the labeling results into the smart contract vulnerability detection model, and obtaining detection results; a classification module for classifying the detection results and completing vulnerability detection; the data preprocessing module, the feature extraction module, and the classification module are connected in sequence.

[0046] The data preprocessing module includes: a contract slicing submodule, which is used to delete useless information from the smart contract and obtain smart contract slices related to the formation of vulnerabilities; a labeling submodule, which is used to label the smart contract slices with different vulnerability types and obtain labeling results; among which, the vulnerability types include: reentrancy vulnerabilities, integer overflow vulnerabilities, denial of service vulnerabilities, TxOrigin vulnerabilities, timestamp dependency vulnerabilities and low-level call vulnerabilities; a transmission submodule, which is used to transmit the labeling results to the feature extraction module; the contract slicing submodule, labeling submodule, transmission submodule and feature extraction module are connected in sequence.

[0047] The feature extraction module includes: a slice acquisition submodule, which is used to obtain annotated historical smart contract slices; a detection model construction submodule, which is used to form a historical word embedding matrix for the annotated historical smart contract slices through the CodeBert pre-training model, and input the historical word embedding matrix into the BiGRU model of the attention mechanism for training to obtain a smart contract vulnerability detection model; the slice acquisition submodule and the detection model construction submodule are connected.

[0048] The detection model construction submodule includes: a pre-training model construction unit, which is used to add words that only belong to solidity to the CodeBert pre-training model to obtain a fine-tuned CodeBert model, and input the annotated historical smart contract slices into the fine-tuned CodeBert model to form a historical word embedding matrix; a detection model construction unit, which is used to input the historical word embedding matrix into the BiGRU model of the attention mechanism for training to obtain a smart contract vulnerability detection model; the pre-training model construction unit and the detection model construction unit are connected.

[0049] The feature extraction module also includes: a detection result acquisition submodule, which is used to input the annotation results into the fine-tuned CodeBert model to form a word embedding matrix, and input the word embedding matrix into the smart contract vulnerability detection model for feature extraction to obtain a feature vector; the smart contract vulnerability detection model includes: a Bi-GRU encoding part and an attention mechanism part; the detection model construction submodule is connected to the detection result acquisition submodule.

[0050] The feature extraction module also includes: a word embedding matrix acquisition unit, which is used to input the annotation results into the fine-tuned CodeBert model to form a word embedding matrix; a hidden layer vector acquisition unit, which is used to obtain a text sequence through the word embedding matrix, input the text sequence into the Bi-GRU encoding part, obtain the output result of the forward hidden layer and the output result of the backward hidden layer, splice the output result of the forward hidden layer and the output result of the backward hidden layer to obtain a hidden layer vector; a feature vector acquisition unit, which is used to input the hidden layer vector into the attention mechanism part, configure the weights of the words in the hidden layer vector according to different importance levels, and combine them to obtain a feature vector; the word embedding matrix acquisition unit, the hidden layer vector acquisition unit and the feature vector acquisition unit are connected in sequence.

[0051] The classification module includes: a classification submodule, which is used to calculate the probability score of each label in the feature vector by using softmax as the activation function and the cross entropy information function as its loss function to obtain the classification result.

[0052] The present invention also discloses a vulnerability detection method for smart contracts, including: collecting smart contracts, deleting useless information from the smart contracts and marking different vulnerability types, and obtaining marking results; building a smart contract vulnerability detection model, inputting the marking results into the smart contract vulnerability detection model, obtaining detection results, classifying the detection results, and completing vulnerability detection.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] As shown in Figure 1, the present invention proposes a smart contract vulnerability detection system based on deep learning, which aims to solve the problem of low accuracy of vulnerability detection on smart contracts.

[0055] The main idea of ​​this invention is to extract the word embedding matrix representation of smart contract slices through CodeBert, and use BiGRU combined with the attention mechanism to extract the sequence features of the smart contract slice sequence to obtain sequence features, thereby solving the problems of insufficient feature extraction ability and low classification accuracy in existing detection methods.

[0056] The technical solutions adopted in the present invention are as follows:

[0057] The deep learning-based smart contract vulnerability detection system proposed in this paper consists of three modules: data preprocessing module, feature extraction module, and classification module. The relationship between these modules follows the following rules:

[0058] (1) The data preprocessing module is responsible for deleting useless information from the collected smart contracts and generating smart contract slices that only contain information related to vulnerability formation, as shown in Figure 2. The smart contract slices are labeled with different categories, such as reentrancy vulnerabilities, integer overflow vulnerabilities, and denial of service vulnerabilities. The smart contract slice results generated by the data preprocessing module will be sent to the feature extraction module.

[0059] (2) The feature extraction module receives the smart contract slices generated by the data preprocessing module. First, the module uses the CodeBert pre-trained model to form a word embedding matrix. Then, the word embedding matrix is ​​sent to the BiGRU model with the attention mechanism for training to obtain the final smart contract vulnerability detection model.

[0060] (3) The classification module classifies the output results of the final smart contract vulnerability detection model.

[0061] The code slice generation process in the aforementioned data preprocessing module includes the following steps: generating a data flow graph, generating a control flow graph, extracting dependencies in the program and generating a program dependency graph, generating different smart contract slices according to different vulnerabilities, and labeling the smart contract slices by category.

[0062] The specific implementation of the aforementioned feature extraction module is as follows:

[0063] ① Using the CodeBert pre-trained model, we generate a word embedding matrix for the smart contract slice sequence obtained in S1. Since the CodeBert pre-trained model does not include solidity, we added words that only belong to solidity to the corpus to allow them to participate in the training. We then use the fine-tuned CodeBert model to generate the word embedding matrix.

[0064] ② Input the word embedding matrix generated by ① into the BiGRU neural network model, and extract the features of the word embedding matrix in combination with the attention mechanism to obtain the feature vector Vec.

[0065] After feature learning is completed, the feature vector representation Vec of the code slice of the smart contract can be obtained. This vector is input into the single-layer connection layer. Through the fully connected layer, the features extracted by Bi-GRU-Att are mapped to K real numbers. These real numbers are determined by the number of categories defined by the multi-classification task.

[0066] ③ Using softmax as the activation function and the cross-entropy information function as its loss function, the probability score of each label is calculated, and the classification result is finally obtained. That is, through the softmax function, these real numbers can be converted into K values ​​between 0 and 1, thus reflecting the probability of each class. Finally, the cross-entropy information function is used as the loss function for training to obtain the model parameters.

[0067] Furthermore, the following six types of vulnerabilities were found in smart contract slices: reentrancy, integer overflow, denial of service, TxOrigin, timestamp dependency, and low-level call. The CodeBert pre-trained model was used to convert these smart contract slices into word embedding matrices, which were then divided into training and test sets in an 8:2 ratio.

[0068] The specific details of the prediction results in ② are as follows:

[0069] Ⅰ. Bi-GRU encoding part: For a given text sequence (word vector matrix) of length L (L is the number of words) that has been converted into code slices, the vector of each word is specified as x tRepresents t∈[1,L], then using the Bi-GRU network, we can get the output of the forward hidden layer at time t and the output of the backward hidden layer As shown in formulas 1-1 and 1-2:

[0070] in, is the output of the forward hidden layer, is the output result of the backward hidden layer, is the forward GRU unit, x t is the vector of the word, is the forward and backward GRU unit output results at time t-1, is the backward GRU unit.

[0071] II. Attention Mechanism: Acquisition and Finally, the attention layer is used to assign weights to the words in the hidden layer of different importance. In the detection task of this article, the words that appear in the calls of custom functions and methods are usually more important than the keywords that already exist in the language. Through formulas (1-3)-(1-5), we can assign different weights to different words, and combine them to form a feature vector representation that can ultimately represent a code sequence.

[0072] In the above formula, W v , b v and m w It is a dynamic parameter and needs to be obtained through training. is the hidden layer vector obtained by concatenating the forward hidden layer and the backward hidden layer, m t is the hidden layer vector mapping result, tanh() is the activation function, for each position i, we calculate an attention weight vector t, softmax() is the loss function, m t T is transposed, L is the total length, vec is the weighted hidden state, t is the current moment, and the attention weight vector is used to calculate the hidden state sequence. Perform weighted summation to obtain a weighted hidden state representation vec, where ak is an example of at.

[0073] Through linear transformation, we map the concatenated hidden layer vector to a new representation m t Then, for each position i, that is, m t For each letter position in , we calculate an attention weight vector at , which is obtained from all positions of m t and a learnable vector m w To calculate it, the softmax function normalizes the value of the vector into a probability distribution.

[0074] Finally, use the attention weight vector a t For the hidden state sequence Perform weighted summation to obtain a weighted hidden state representation, which is vec in the above formula. The purpose of this step is to combine the attention weight with the hidden state to extract the most relevant information in the input sequence.

[0075] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A vulnerability detection system for smart contracts, characterized in that: include: A data preprocessing module is used to collect smart contracts, delete useless information from the smart contracts, obtain smart contract slices related to the formation of vulnerabilities, mark the smart contract slices with different vulnerability types, and obtain marking results; A feature extraction module, used to build a smart contract vulnerability detection model based on the annotation results, and obtain detection results based on the smart contract vulnerability detection model; A classification module, used to classify the detection results and complete vulnerability detection; The data preprocessing module, the feature extraction module and the classification module are connected in sequence.

2. The vulnerability detection system for smart contracts according to claim 1, characterized in that: The data preprocessing module comprises: The contract slicing submodule is used to delete useless information from the smart contract and obtain smart contract slices related to the formation of the vulnerability; A labeling submodule is used to label the smart contract slice with different vulnerability types and obtain the labeling results; wherein the vulnerability types include: reentrancy vulnerability, integer overflow vulnerability, denial of service vulnerability, TxOrigin vulnerability, timestamp dependency vulnerability and low-level call vulnerability; The transmission submodule is used to transmit the annotation result to the feature extraction module.

3. The vulnerability detection system for smart contracts according to claim 1, characterized in that: The feature extraction module comprises: The slice acquisition submodule is used to obtain the annotated historical smart contract slices; The detection model construction submodule is used to form a historical word embedding matrix for the annotated historical smart contract slices through the CodeBert pre-training model, input the historical word embedding matrix into the BiGRU model of the attention mechanism for training, and obtain the smart contract vulnerability detection model.

4. The vulnerability detection system for smart contracts according to claim 3 is characterized in that: The detection model building submodule includes: A pre-trained model construction unit, used for adding words belonging only to solidity to the CodeBert pre-trained model, obtaining a fine-tuned CodeBert model, and inputting the annotated historical smart contract slices into the fine-tuned CodeBert model to form the historical word embedding matrix; The detection model building unit is used to input the historical word embedding matrix into the BiGRU model of the attention mechanism for training to obtain the smart contract vulnerability detection model.

5. The vulnerability detection system for smart contracts according to claim 4, characterized in that: The feature extraction module also includes: The detection result acquisition submodule is used to input the annotation results into the fine-tuned CodeBert model to form the word embedding matrix, and input the word embedding matrix into the smart contract vulnerability detection model for feature extraction to obtain a feature vector; wherein the smart contract vulnerability detection model includes: a Bi-GRU encoding part and an attention mechanism part.

6. The vulnerability detection system for smart contracts according to claim 5, characterized in that: The feature extraction module also includes: A word embedding matrix acquisition unit, used for inputting the annotation result into the fine-tuned CodeBert model to form the word embedding matrix; A hidden layer vector acquisition unit, used for inputting the word embedding matrix into the Bi-GRU encoding part, acquiring the output result of the forward hidden layer and the output result of the backward hidden layer, and concatenating the output result of the forward hidden layer and the output result of the backward hidden layer to acquire a hidden layer vector; The feature vector acquisition unit is used to input the hidden layer vector into the attention mechanism part, configure the weights of the words in the hidden layer vector according to different importance levels, and combine them to obtain the feature vector.

7. The vulnerability detection system for smart contracts according to claim 6, characterized in that: The method to obtain the output results of the forward hidden layer and the output results of the backward hidden layer is: in, is the output result of the forward hidden layer, is the output result of the backward hidden layer, is the forward GRU unit, x t is the vector of the word, They are the forward and backward GRU unit output results at time t-1, is the backward GRU unit.

8. The vulnerability detection system for smart contracts according to claim 6, characterized in that: The method for obtaining the feature vector is: Among them, m t is the hidden layer vector mapping result, W v 、m w , b v These are dynamic parameters. is the hidden layer vector, tanh() is the activation function, a t is the attention weight vector, softmax() is the loss function, m t T is the transpose, a k for a t , L is the total length, vec is the weighted hidden state, and t is the current moment.

9. The vulnerability detection system for smart contracts according to claim 6, characterized in that: The classification module comprises: The classification submodule is used to calculate the probability score of each label in the feature vector by using softmax as the activation function and the cross entropy information function as its loss function to obtain the classification result.

10. A vulnerability detection method for smart contracts, characterized in that: include: Collect smart contracts, delete useless information and mark different vulnerability types of the smart contracts, and obtain marking results; Construct a smart contract vulnerability detection model, input the annotation results into the smart contract vulnerability detection model, obtain detection results, classify the detection results, and complete vulnerability detection.

Images