Line-level software vulnerability detection method and detection system fused with context semantics

Abstract

The invention relates to a row-level software vulnerability detection method and detection system fusing context semantics. The row-level software vulnerability detection method fused with context semantics comprises the following steps: S1, preprocessing a target software code to remove noise to obtain a preprocessed software code; s2, carrying out vector embedding on the nth row of codes to obtain an nth code row vector Ln; and associating the position information of the code line for the nth code line vector Ln. According to the method, each row of codes are converted into vector representation, then context information fusion is carried out on the code rows in the context information extraction and feature fusion stages, so that global information related to the code rows is captured, and all context information and code row information are fused immediately; whether the code lines have vulnerabilities or not is judged by judging the weights or probabilities occupied by the codes of all the lines, the line-level vulnerability detection task of the codes is finally completed, and the code vulnerability positioning precision is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of software vulnerability detection, and in particular to a row-level software vulnerability detection method and a detection system integrating context semantics. Background Art

[0002] Nowadays, many industries use software as a carrier of their business, and the number of software has increased dramatically. However, in the process of software development, due to inevitable human factors and insufficient security testing, there are a large number of vulnerabilities in the software. Vulnerabilities in the software may be exploited by malicious elements, which means they can remotely execute code to achieve data theft, surveillance activities and malware propagation. Therefore, it is crucial to detect code vulnerabilities during the software development process.

[0003] Traditional vulnerability detection often consumes a lot of energy and resources, and is heavily dependent on the subjective experience and knowledge reserves of domain experts. Modern software systems have a variety of vulnerabilities, which are very complex and difficult to detect through specific methods. Most of the existing software vulnerability detection methods that apply deep learning focus on coarse-grained levels, such as slice-level vulnerability detection, function-level vulnerability detection, and file-level vulnerability detection. However, not all lines of a vulnerable function are defective (approximately 0.03%-2.9%). Therefore, after applying the method, developers still need a lot of energy to locate the specific vulnerable lines of code. Summary of the invention

[0004] Based on this, it is necessary to provide a row-level software vulnerability detection method and detection system that integrates contextual semantics to address the problem of insufficient vulnerability location accuracy in existing detection methods. It uses Transformer to fuse context information, and uses compressed bilinear pooling technology to fuse context information and code line information to ultimately complete the vulnerability detection task.

[0005] In a first aspect, the present invention proposes a row-level software vulnerability detection method integrating context semantics, which is used to perform row-level vulnerability detection on target software code; the row-level software vulnerability detection method comprises the following steps:

[0006] S1, preprocessing the target software code to remove noise, and obtaining the preprocessed software code;

[0007] The preprocessed software code includes N lines of code;

[0008] S2. Embed the nth line of code into a vector to obtain the nth line of code vector L n , n∈[1,N];

[0009] For the nth code line vector L nThe location information of the associated code line;

[0010] S3, transform the nth code line vector L n The nth context vector LC is calculated by the encoding module n ;

[0011] S4, through the weight matrix U for the nth code line vector L n Learn and get the nth code line vector L′ n ;

[0012] The nth context vector LC is calculated by the weight matrix V n Learn and get the nth context vector LC′ n ;

[0013] S5. Convert N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n}Perform feature fusion and calculate the attention mapping matrix Attention_map through the projection model;

[0014] S6. Extract vulnerability scores from the attention mapping matrix Attention_map through matrix operations to obtain N vulnerability scores.

[0015] Arrange the N vulnerability scores from large to small;

[0016] The nth vulnerability score corresponds to the nth code line vector L n ;

[0017] The nth code line vector L n According to the position information, the nth line of code in the software code is corresponded;

[0018] Among them, the code lines corresponding to the vulnerability scores in the first M positions are judged as: having vulnerabilities;

[0019] M is a predefined parameter.

[0020] As a preferred example, in S2, the nth code line vector L is quantized using sine and cosine functions. n The location information of the associated code line is as follows:

[0021]

[0022]

[0023] Where P is the position information of the code line, i is the position of the code line, k is the dimension index in the position encoding, and d is the dimension of the code line vector L.

[0024] As a preferred example, in S3, the encoding module includes an encoder in a Transformer model.

[0025] As a preferred example, in S3, the context vector LC is calculated. n Previously, the number of code lines in all target software codes in the same batch was filled to K; K was the maximum number of code lines in all target software codes in the batch.

[0026] As a preferred example, in S4, the code line vector L n The formula for learning is:

[0027] L′ n =ReLU(L n *U)

[0028] The context vector LC n The formula for learning is:

[0029] LC′ n =ReLU(LC n *V)

[0030] Where ReLU is a nonlinear activation function.

[0031] As a preferred example, in S5, the method for performing feature fusion includes:

[0032] Generate a projection of a low-rank space using projection matrix P and projection matrix Q;

[0033] Use projection to perform attention mapping to obtain the attention mapping matrix Attention_map;

[0034] The calculation formula of the attention mapping matrix Attention_map is:

[0035] Attention_map=(P·{L′1,L′2,…,L′ n})·(Q·{LC′1,LC′2,…,LC′ n}).

[0036] As a preferred example, in S5, the diagonal elements in the attention mapping matrix Attention_map are extracted to obtain a new mapping matrix;

[0037] Normalize the elements of each row in the new mapping matrix;

[0038] Each element in the normalized mapping matrix corresponds to a vulnerability score.

[0039] As a preferred example, the row-level software vulnerability detection method integrating context semantics also includes S0;

[0040] S0, performing a coarse-grained detection on the target software code to detect whether there are vulnerabilities in the target software code as a whole;

[0041] If it is detected that the target software code as a whole has vulnerabilities, then proceed to S1-S6;

[0042] If it is detected that there are no vulnerabilities in the target software code as a whole, the process ends.

[0043] As a preferred example, the method for performing coarse-grained detection on the target software code includes the following steps:

[0044] S21, preprocessing the target software code to remove noise, and obtaining a preprocessed software code;

[0045] The preprocessed software code includes N lines of code;

[0046] S22, embed the nth line of code into a vector to obtain the nth line of code vector L n , n∈[1,N];

[0047] For the nth code line vector L n Associated location information;

[0048] S23, the nth code line vector L n The nth context vector LC is calculated by the encoding module n ;

[0049] S24, the nth code line vector L is calculated by the weight matrix U n Learn and get the nth code line vector L′ n ;

[0050] The nth context vector LC is calculated by the weight matrix V n Learn and get the nth context vector LC′ n ;

[0051] S25, transform N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n}Perform feature fusion to obtain the attention mapping matrix Attention_map;

[0052] S26. Pool the attention map matrix Attention_map and input it into the linear layer to calculate the variable value B. The calculation formula is:

[0053] B=Sigmoid(Linear(mean(Attention_map))

[0054] In the formula, Sigmoid is the Sigmoid activation function, and Linear represents linear processing;

[0055] Map the variable value B to between 0 and 1. A value greater than or equal to 0.5 after mapping is considered to have a vulnerability, and a value less than 0.5 is considered to have no vulnerability.

[0056] In a second aspect, the present invention proposes a row-level software vulnerability detection system integrating context semantics, which uses the row-level software vulnerability detection method integrating context semantics in the first aspect; the row-level software vulnerability detection system comprises:

[0057] A preprocessing module, which is used to preprocess the target software code to remove noise;

[0058] The vector embedding module is used to embed the nth line of code into a vector to obtain the nth line of code vector L n ; It is also used to integrate the code line location information into the code line vector L n middle;

[0059] The information extraction module is used to fuse the context information according to the nth code line vector L n Calculate the nth context vector LC n ; Also used for the nth code line vector L n and the nth context vector LC n Perform learning and obtain the learned code row vector L′ n and context vector LC′ n ;

[0060] The vulnerability prediction module is used to transform N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n The projection matrix P and the projection matrix Q are used to generate a projection of a low-rank space, and the attention mapping matrix Attention_map is calculated. The attention mapping matrix Attention_map is also used to extract vulnerability scores through matrix operations to obtain N vulnerability scores. The N vulnerability scores are arranged from large to small, and the code lines corresponding to the vulnerability scores in the first M positions are judged to have vulnerabilities.

[0061] The beneficial effect of the present invention is that the present invention converts each line of code into a vector representation, and then fuses the context information of the code line in the context information extraction and feature fusion stage, thereby capturing the global information related to the code line, and then fuses all the context information and code line information, and judges whether the code line has a vulnerability by judging the weight or probability size (i.e., vulnerability score) of each line of code, and finally completes the line-level vulnerability detection task of the code, greatly improving the accuracy of code vulnerability positioning, and reducing the effort spent by developers to locate specific lines of code with vulnerabilities. In addition, the present invention realizes automation of the entire process from code preprocessing to vulnerability prediction, providing developers with an efficient vulnerability positioning tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 creative labor.

[0063] Figure 1 This is a flow chart of the row-level software vulnerability detection method proposed by the present invention;

[0064] Figure 2 This is an example diagram showing the comparison before and after the code in one of the embodiments is preprocessed;

[0065] Figure 3 is a flow chart of a method for coarse-grained software vulnerability detection in another embodiment;

[0066] Figure 4 The figure shows the result of vulnerability detection using the row-level software vulnerability detection method. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "or / and" used herein includes any and all combinations of one or more of the related listed items.

[0069] This embodiment provides a line-level software vulnerability detection method that integrates contextual semantics, which mainly includes the steps of code preprocessing, code line vector embedding, context information extraction, feature fusion and vulnerability prediction. Code preprocessing first removes the noise caused by personalized naming and comments, blank lines in the code, special characters, etc. In the code line vector embedding stage, each line of code is converted into a vector representation through tools such as Code Tokenizer and CodeBERT model. In the context information extraction and feature fusion stages, Transformer is used to fuse context information to capture global information related to the code line, and then compressed bilinear pooling technology is used to fuse the context information and code line information to finally complete the vulnerability detection task.

[0070] For details, please refer to Figure 1 ,The row-level software vulnerability detection method integrating context semantics includes the following steps:

[0071] S1. Preprocess the target software code to remove noise and obtain preprocessed software code. The preprocessed software code includes N lines of code.

[0072] In this step, the line-level software vulnerability detection method can perform vulnerability detection on the code in batches. Each batch includes multiple files. Each file includes multiple lines of code. Vulnerability detection tasks often ignore the different naming styles of functions, variables, methods, etc. and the unnecessary noise introduced by comments in the code. In this embodiment, in order to reduce the impact of such noise, the code preprocessing stage first replaces the personalized names in all input source domain and target domain function codes with symbolic names, for example, Figure 2 As shown in the figure, change the variable names to "VAR0" and "VAR1", the function names to "FUN0" and "FUN1", and delete the comments, blank lines, special characters, etc. contained in the code.

[0073] S2. Embed the nth line of code into a vector to obtain the nth line of code vector L n , n∈[1,N]. For the nth code line vector L n Associated location information.

[0074] The preprocessed software code includes N lines of code. Code Tokenizer is used to perform word segmentation and ID conversion on each line of code. Then, the ID corresponding to each line of code is converted into a vector representation through the CodeBERT model to obtain multiple code line vectors L n, n∈[1,N], n is an integer between 1 and N. The CodeBERT model is based on the Transformer architecture and can effectively handle tasks related to understanding and generating code. The CodeBERT model can effectively capture meaningful semantic information of code lines, which helps to construct a vector representation of code lines. The specific operation is to convert each line of code obtained in the code preprocessing stage into a corresponding vector representation through the pre-trained CodeBERT model. In this embodiment, the code line vector L n The dimension of is set to d = 768, that is, each line of code in a function is converted into a vector of 768 dimensions. The code line vector L input in the subsequent model n It is expressed in code form as (number of code lines, embedding vector dimension). In this embodiment, vulnerability detection can be performed on software code in batches. Each batch includes multiple files, and the target software code in each file includes several lines of code.

[0075] To ensure uniform dimensionality, in the code line vector L n Before subsequent processing, fill the number of lines of code in each file in each batch to the maximum number of lines of code in this batch. For example, the input shape is (number of batches, number of lines of code in each file, 768) n , by filling the vector with 0, it is transformed into a code line vector L with a shape of (number of batches, maximum number of code lines in this batch file, 768) n In addition, the sine-cosine function is used to associate the location information of the code line with the code line vector L n The associated formula is:

[0076]

[0077] Where P is the position information of the code line. i is the position of the code line, ranging from 0 to the maximum number of code lines - 1. k is the dimension index in the position encoding, ranging from 0 to d / 2-1. d is the code line vector L n Dimension.

[0078] S3, transform the nth code line vector L n The nth context vector LC is calculated by the encoding module n .

[0079] In this step, the encoding module can use the encoder in the Transformer model, whose main function is to convert the input sequence into a feature representation (combining its context information). According to the task requirements, the encoder part in the Transformer model is used here to process the input sequence and generate a context representation. The encoding module in this embodiment can be composed of two of the above encoders connected in series. The nth code line vector L after the above processingn The nth context vector LC that integrates its context information is calculated in the input encoder n .

[0080] S4, through the weight matrix U for the nth code line vector L n Learn and get the nth code line vector L′ n . The nth context vector LC is transformed by the weight matrix V n Learn and get the nth context vector LC′ n .

[0081] Specifically, we first define two learning weight matrices U and V, and then perform the following operations on N code line vectors {L1, L2, ...L n} and N context vectors {LC1, LC2, …LC n}For further study, the learning formula is:

[0082] L′ n =ReLU(L n *U)

[0083] LC′ n =ReLU(LC n *V)

[0084] In the formula, ReLU is a nonlinear activation function, the purpose is to add certain nonlinear factors. L′ n , LC′ n Represented as the learned code row vector L n and context vector LC n .

[0085] S5. Convert N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n}Perform feature fusion to obtain the attention mapping matrix Attention_map.

[0086] In order to obtain a more accurate representation, this method uses compressed bilinear pooling technology to perform code line vector L′ n and the corresponding context vector LC′ n to be processed.

[0087] The N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n}Use the projection matrix P and the projection matrix Q to generate a projection of a low-rank space, and use the MCB (multimodal compressed bilinear pooling model) to calculate the attention mapping matrix Attention_map, which is calculated as follows:

[0088] Attention_map=(P·{L′1,L′2,…,L′ n})·(Q·{LC′1,LC′2,…,LC′ n});

[0089] In this embodiment, the shapes of the projection matrix P and the projection matrix Q can both be expressed as [1, h_out, rank, hidden_dim*k]. h_out represents the number of output heads, rank is the rank of the low-rank matrix, hidden_dim is the hidden layer dimension, and k is the parameter of low-rank pooling. In this embodiment, the parameters can be specifically set to h_out=2, rank=10, hidden_dim=256, k=3. The purpose of the above-mentioned low-rank projection is to map a high-rank matrix to a space with a lower rank through projection, which can effectively reduce the complexity of the data, while retaining the main features of the data and performing feature fusion.

[0090] S6. Extract the vulnerability scores from the attention mapping matrix through matrix operations to obtain N vulnerability scores. Arrange the N vulnerability scores from large to small. The nth vulnerability score corresponds to the nth code line vector L n The nth code line vector L n According to the location information, the nth line of code in the software code is corresponding. Among them, the code line corresponding to the vulnerability score in the first M positions is judged as having a vulnerability. M is a predefined parameter.

[0091] In this step, the attention mapping matrix Attention_map is subjected to matrix operations to obtain the vulnerability score corresponding to each line of code. The vulnerability scores are arranged from large to small, and the code lines corresponding to the vulnerability scores within the set sequence number (i.e., the first several numbers in the set sorting) are judged to have vulnerabilities. Specifically, after the above model learning, the attention mapping matrix Attention_map with the shape of (number of batches, number of output heads, maximum number of code lines in this batch function, maximum number of code lines in this batch function) is obtained. In this embodiment, the number of output heads is 2, so the attention scores obtained in the two subspaces corresponding to the number of output heads are summed to obtain a further (number of batches, maximum number of code lines in this batch function, maximum number of code lines in this batch function) attention mapping matrix 2. Then take the diagonal elements in the attention mapping matrix 2 to obtain a new attention mapping matrix 3 (used to extract the attention of each element to itself from the original attention mapping matrix) of (number of batches, maximum number of code lines in this batch function). Add the values ​​of each row in the attention mapping matrix three and divide the elements of each row in the attention mapping matrix three by the sum of the corresponding number of rows, which is equivalent to making the sum of the weights of each row in the attention mapping matrix three equal to 1 and normalizing the elements therein. Then extract the elements in the matrix to obtain N vulnerability scores containing the location information of the code lines. Arrange the N vulnerability scores from large to small, and the code lines corresponding to the vulnerability scores in the first M positions are judged to have vulnerabilities. M is a predefined parameter, so that the code lines most likely to have vulnerabilities can be located.

[0092] In order to verify the detection effect of the line-level software vulnerability detection method integrating context semantics, the line-level software vulnerability detection method integrating context semantics in the above embodiment is used to perform vulnerability detection on a set of target software codes. The target software codes are known to have vulnerabilities through running in advance, but it is unknown which specific line of code has the vulnerability. Figure 3 As shown, through the vulnerability detection of the line-level software vulnerability detection method, a probability (i.e., the vulnerability score calculated above) can be calculated for each line of code, and the code lines with vulnerabilities can be determined by sorting the probabilities, such as Figure 3 The 4th / 5th lines of code in the , which are ranked at the top in probability, can be identified as having vulnerabilities. For the detected lines of code, after manual inspection:

[0093] Line 4: Array out of bounds exception. The loop condition is written as i <= input.length(). When i = input.length(), input.charAt(i) will throw a StringIndexOutOfBoundsException.

[0094] Line 5: NullPointerException. If input is null, calling input.charAt(i) will throw a NullPointerException.

[0095] From this we can see that Figure 3 There are vulnerabilities in lines 4 and 5 of the program, and the line-level software vulnerability detection method that integrates context semantics can accurately detect the code lines where the vulnerabilities are located. This line-level software vulnerability detection method can locate the specific code lines with vulnerabilities, greatly improving the vulnerability location accuracy of the software code.

[0096] In another embodiment, the row-level software vulnerability detection method integrating context semantics also includes S0.

[0097] S0. Perform a coarse-grained detection on the target software code to detect whether there are vulnerabilities in the target software code as a whole.

[0098] If it is detected that the target software code as a whole has vulnerabilities, then proceed to S1-S6.

[0099] If it is detected that there are no vulnerabilities in the target software code as a whole, the process ends.

[0100] In this row-level software vulnerability detection method that integrates contextual semantics, S1-S6 can perform fine-grained vulnerability detection, while S0 can perform coarse-grained file-level vulnerability detection. Specifically, Figure 4 As shown, the method for performing coarse-grained detection on the target software code includes the following steps:

[0101] S21, preprocessing the target software code to remove noise, and obtaining a preprocessed software code;

[0102] The preprocessed software code includes N lines of code;

[0103] S22, embed the nth line of code into a vector to obtain the nth line of code vector L n , n∈[1,N];

[0104] For the nth code line vector L n Associated location information;

[0105] S23, the nth code line vector L n The nth context vector LC is calculated by the encoding module n ;

[0106] S24, the nth code line vector L is calculated by the weight matrix U n Learn and get the nth code line vector L′ n ;

[0107] The nth context vector LC is calculated by the weight matrix V n Learn and get the nth context vector LC′ n ;

[0108] S25, transform N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n}Perform feature fusion to obtain the attention mapping matrix Attention_map;

[0109] S26. Perform an average pooling operation on the attention mapping matrix Attention_map, and input the pooled result into the linear layer (a fully connected layer can be used: Fully Connected Layer) to calculate the variable value B. The calculation formula is:

[0110] B=Sigmoid(Linear(mean(Attention_map))

[0111] Where Sigmoid is the Sigmoid activation function, and Linear represents linear processing.

[0112] The obtained variable value is mapped to between 0 and 1, and a binary classification task is performed: a mapped value greater than or equal to 0.5 is considered to have a vulnerability, and the system output value is 1. A mapped value less than 0.5 is considered to have no vulnerability, and the system output value is 0.

[0113] This coarse-grained vulnerability detection method can perform coarse-grained detection on the target software code before performing line-level vulnerability detection on the target software code to detect whether the target software as a whole has vulnerabilities. Fine-grained vulnerability detection is performed after determining that the target code has vulnerabilities, thereby reducing the waste of computing resources. On the other hand, the file-level vulnerability detection results and data can be used as part of the data for pre-training or training of each model in the fine-grained vulnerability detection method, increasing the amount of data and improving the efficiency of model training.

[0114] In addition, the model used in the row-level software vulnerability detection method that integrates contextual semantics can also be trained using existing methods: supervised training of the model using training sets and test sets, and updating the model parameters through back propagation until the training loss function reaches the expected value.

[0115] In another embodiment, a row-level software vulnerability detection system integrating context semantics is also proposed, which uses the row-level software vulnerability detection method integrating context semantics as described above. The row-level software vulnerability detection system includes a preprocessing module, a vector embedding module, an information extraction module, and a vulnerability prediction module.

[0116] The preprocessing module is used to preprocess the target software code to remove noise.

[0117] The vector embedding module is used to embed the nth line of code into a vector to obtain the nth line of code vector L n The vector embedding module is also used to incorporate the code line location information into the code line vector L n middle.

[0118] The information extraction module is used to fuse context information and extract the nth code line vector L n Calculate the nth context vector LC n The information extraction module is also used to extract the nth code line vector L n and the nth context vector LC n Perform learning and obtain the learned code row vector L′ n and context vector LC′ n .

[0119] The vulnerability prediction module is used to transform N code line vectors {L′1, L′2, …, L′ n} and N context vectors {LC′1, LC′2, …, LC′ n}Use the projection matrix P and the projection matrix Q to generate a projection of a low-rank space, and calculate the attention mapping matrix Attention_map. The vulnerability prediction module is also used to extract the vulnerability score of the attention mapping matrix Attention_map through matrix operations to obtain N vulnerability scores. Then the N vulnerability scores are arranged from large to small, and the code lines corresponding to the vulnerability scores in the first M positions are judged to have vulnerabilities.

[0120] In another embodiment, a row-level software vulnerability detection system integrating contextual semantics is also proposed. Compared with the above-mentioned row-level software vulnerability detection system integrating contextual semantics, the difference is that the vulnerability prediction module is also used to input the attention mapping matrix Attention_map into the linear layer after pooling to obtain the variable value B, and map the variable value B to between 0 and 1. The mapped value is greater than or equal to 0.5, which is considered to have a vulnerability, and less than 0.5, which is considered to have no vulnerability.

[0121] In some other embodiments, an electronic device is also provided. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of the above-mentioned row-level software vulnerability detection method integrating context semantics are implemented.

[0122] Some other embodiments also provide a computer-readable storage medium storing a computer program that implements the above-mentioned steps of the row-level software vulnerability detection method integrating context semantics when the computer program is executed by a processor.

[0123] Some other embodiments also provide a software program product, which includes program instructions, and when the software program product is run on an electronic device, the electronic device executes the above-mentioned row-level software vulnerability detection method integrating context semantics.

[0124] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A row-level software vulnerability detection method integrating contextual semantics, characterized in that , which is used to perform line-level vulnerability detection on the target software code; the line-level software vulnerability detection method includes the following steps: S1, preprocessing the target software code to remove noise, and obtaining the preprocessed software code; The preprocessed software code includes N lines of code; S2. Perform vector embedding on the nth line of code in the preprocessed software code to obtain the nth code line vector L n , n∈[1,N]; For the nth code line vector L n The location information of the associated code line; S3, transform the nth code line vector L n The nth context vector LC is calculated by the encoding module n ; S4, through the weight matrix U for the nth code line vector L n Learn and get the nth code line vector L ′ n ; The nth context vector LC is calculated by the weight matrix V n Learn and get the nth context vector LC ′ n ; S5. Transform N code lines into vectors {L ′ 1.L ′ 2, ..., L ′ n } and N context vectors {LC ′ 1.LC ′ 2, ..., LC ′ n }Perform feature fusion to obtain the attention mapping matrix Attention_map; S6. Extract vulnerability scores from the attention mapping matrix Attention_map through matrix operations to obtain N vulnerability scores. Arrange the N vulnerability scores from large to small; The nth vulnerability score corresponds to the nth code line vector L n ; The nth code line vector L n According to the position information, the nth line of code in the software code is corresponded; Among them, the code lines corresponding to the vulnerability scores in the first M positions are judged as: having vulnerabilities; M is a predefined parameter.

2. The row-level software vulnerability detection method integrating context semantics according to claim 1 is characterized in that: In S2, the nth code line vector L is calculated using sine and cosine functions. n The location information of the associated code line is as follows: Where P is the position information of the code line, i is the position of the code line, k is the dimension index in the position encoding, and d is the dimension of the code line vector L.

3. The row-level software vulnerability detection method integrating context semantics according to claim 1 is characterized in that: In S3, the encoding module includes the encoder in the Transformer model.

4. The row-level software vulnerability detection method integrating contextual semantics according to claim 1 is characterized in that: In S3, the context vector LC is calculated n Previously, the number of code lines in all target software codes in the same batch was filled to K; K was the maximum number of code lines in all target software codes in the batch.

5. The row-level software vulnerability detection method integrating context semantics according to claim 1 is characterized in that: In S4, the code line vector L n The formula for learning is: L ′ n =ReLU(L n *U) The context vector LC n The formula for learning is: IC ′ n =ReLU(LC n *V) In the formula, ReLU is a nonlinear activation function.

6. The row-level software vulnerability detection method integrating context semantics according to claim 1 is characterized in that: In S5, the method for feature fusion includes: Generate a projection of a low-rank space using projection matrix P and projection matrix Q; Use projection to perform attention mapping to obtain the attention mapping matrix Attention_map; The calculation formula of the attention mapping matrix Attention_map is: Attention_map=(P·{L ′ 1,L ′ 2,…,L ′ n })·(Q·{LC ′ 1,LC ' 2,…,LC ' n })。 7. The row-level software vulnerability detection method integrating context semantics according to claim 1 is characterized in that: In S6, the steps of performing matrix operations include: Extract the diagonal elements in the attention mapping matrix Attention_map to obtain a new mapping matrix; Normalize the elements of each row in the new mapping matrix; Each element in the normalized mapping matrix corresponds to a vulnerability score.

8. The row-level software vulnerability detection method integrating context semantics according to claim 7 is characterized in that: The row-level software vulnerability detection method integrating context semantics also includes S0; S0, performing a coarse-grained detection on the target software code to detect whether there are vulnerabilities in the target software code as a whole; If it is detected that the target software code as a whole has vulnerabilities, then proceed to S1-S6; If it is detected that there are no vulnerabilities in the target software code as a whole, the process ends.

9. The row-level software vulnerability detection method integrating context semantics according to claim 8 is characterized in that: The method for coarse-grained detection of the target software code includes the following steps: S21, preprocessing the target software code to remove noise, and obtaining a preprocessed software code, wherein the preprocessed software code includes N lines of code; S22, embed the nth line of code into a vector to obtain the nth line of code vector L n , n∈[1,N]; For the nth code line vector L n Associated location information; S23, the nth code line vector L n The nth context vector LC is calculated by the encoding module n ; S24, the nth code line vector L is calculated by the weight matrix U n Learn and get the nth code line vector L ' n ; The nth context vector LC is calculated by the weight matrix V n Learn and get the nth context vector LC ′ n ; S25, transform N code line vectors {L ′ 1.L ' 2, ..., L ′ n } and N context vectors {LC ′ 1.LC ′ 2, ..., LC ′ n }Perform feature fusion to obtain the attention mapping matrix Attention_map; S26. Pool the attention map matrix Attention_map and input it into the linear layer to calculate the variable value B. The calculation formula is: B=Sigmoid(Linear(mean(Attention_map)) In the formula, Sigmoid is the Sigmoid activation function, and Linear represents linear processing; Map the variable value B to between 0 and 1. A value greater than or equal to 0.5 after mapping is considered to have a vulnerability, and a value less than 0.5 is considered to have no vulnerability.

10. A row-level software vulnerability detection system integrating contextual semantics, characterized in that: It uses the row-level software vulnerability detection method integrating contextual semantics as described in any one of claims 1 to 9; The row-level software vulnerability detection system includes: A preprocessing module, which is used to preprocess the target software code to remove noise; The vector embedding module is used to embed the nth line of code into a vector to obtain the nth line of code vector L n ; It is also used to integrate the code line location information into the code line vector L n middle; The information extraction module is used to fuse the context information according to the nth code line vector L n Calculate the nth context vector LC n ; Also used for the nth code line vector L n and the nth context vector LC n Perform learning and obtain the learned code line vector L ′ n and context vector LC ′ n ; Vulnerability prediction module, which is used to transform N code line vectors {L ′ 1.L ′ 2, ..., L ′ n } and N context vectors {LC ′ 1.LC ′ 2, ..., LC ′ n The projection matrix P and the projection matrix Q are used to generate a projection of a low-rank space, and the attention mapping matrix Attention_map is calculated. The attention mapping matrix Attention_map is also used to extract vulnerability scores through matrix operations to obtain N vulnerability scores. The N vulnerability scores are arranged from large to small, and the code lines corresponding to the vulnerability scores in the first M positions are judged to have vulnerabilities.

Images