CN113378176B - Software vulnerability identification method based on graph neural network detection with weight deviation - Google Patents

Software vulnerability identification method based on graph neural network detection with weight deviation Download PDF

Info

Publication number
CN113378176B
CN113378176B CN202110652749.4A CN202110652749A CN113378176B CN 113378176 B CN113378176 B CN 113378176B CN 202110652749 A CN202110652749 A CN 202110652749A CN 113378176 B CN113378176 B CN 113378176B
Authority
CN
China
Prior art keywords
node
information
nodes
graph
time step
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110652749.4A
Other languages
Chinese (zh)
Other versions
CN113378176A (en
Inventor
李辉
曲阳
刘慧江
汪海博
赵娇茹
刘勇
郭世凯
陈荣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian Maritime University
Original Assignee
Dalian Maritime University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dalian Maritime University filed Critical Dalian Maritime University
Priority to CN202110652749.4A priority Critical patent/CN113378176B/en
Publication of CN113378176A publication Critical patent/CN113378176A/en
Application granted granted Critical
Publication of CN113378176B publication Critical patent/CN113378176B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/50—Monitoring users, programs or devices to maintain the integrity of platforms, e.g. of processors, firmware or operating systems
    • G06F21/57—Certifying or maintaining trusted computer platforms, e.g. secure boots or power-downs, version controls, system software checks, secure updates or assessing vulnerabilities
    • G06F21/577—Assessing vulnerabilities and evaluating computer system security
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Hardware Design (AREA)
  • Computer Security & Cryptography (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Computing Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a software vulnerability identification method with weight deviation based on graph neural network detection, which comprises the following steps: will beThe word and the symbol in the source code are regarded as a node, the source code is expressed by adopting a composition graph, and an initial characteristic value of each node and an initial characteristic of a connected edge of each graph are obtained; the generated constitution diagram is taken as input, the information output by the reset gate, the information output by the update gate and the information of the self node are finally combined, and then the node activation value is output through the Sigmoid activation function to be taken as the node state at the final moment; when the word nodes are sufficiently updated, the words are aggregated into a graphical level representation of the function code, and a final vulnerability recognition result is generated based on the representation; using
Figure DDA0003112340680000011
The Dice coefficient minimizes the defective software bug loss value, thereby identifying the software bug.

Description

Software vulnerability identification method based on graph neural network detection with weight deviation
Technical Field
The invention relates to the technical field of software monitoring, in particular to a software vulnerability identification method with weight deviation based on graph neural network detection.
Background
Software vulnerability detection is one of the main means to check and discover the existence of security vulnerabilities in software systems. The design errors, coding defects and operation faults of the software are searched by utilizing various tools to audit codes in the software or analyzing the execution process of the software. Early vulnerability detection techniques are classified into static analysis methods and dynamic analysis methods depending on whether program operation is relied on. As machine learning evolves, researchers have begun to attempt to combine traditional techniques with machine learning to predict code vulnerabilities. The static analysis method is generally applied to a development and coding stage of software, the software is not required to be run, and vulnerabilities are found by scanning information such as source code analysis lexical, grammar, control flow, data flow and the like. The dynamic analysis method is generally applied to a test operation stage of software, and in the process of operating a software program, vulnerabilities are discovered by analyzing information such as the state and execution path of the program in a dynamic debugger. The dynamic analysis method is executed by adopting symbols in the past to carry out vulnerability mining and the dynamic analysis method is adopted to check errors by adopting fuzzy test. Further to the following, researchers consider the source code as a flat sequence (resembling natural language), extracting features of the source code to improve vulnerability recognition. The latest progress of deep learning models in machine learning is to extract structural features of source codes by using Abstract Syntax Tree (AST) Control Flow Graph (CFG), data Flow Graph (DFG) and Natural Code Sequence (NCS), form a composite code representation with AST as backbone, and perform vulnerability recognition by gating a graph loop layer and a convolution conversion layer.
The effective representation of the source code is lacking. Most of the previous methods related to code prediction process source codes according to sequences, which makes it difficult for a model to learn the syntax structure information of the source codes. However, in contrast to weakly structured natural languages, programming languages are formal languages, and source code written with them is explicit and structured. To overcome limitations in code representation, many working views explore more code representation structures, such as tree structures or graph structures. Therefore, it is possible to efficiently represent the syntax structure information of the code by using a specific graph.
There is a lack of focus learning on defective samples. In the past, most of the work uses a cross entropy loss function in terms of loss functions, and the ratio of the loss function weight of a predicted defective sample to a non-defective sample is 1:1. However, in practical applications, defective samples often account for a small number of the total codes. When the weight of the loss function of the defective sample is the same as the proportion of the sample without defects, the model learns that the defective sample is insufficient, and the prediction result is biased to the code without defects. Therefore, consider how to increase the loss function weight of the defective sample, so that the model can effectively learn the characteristics of the defective sample.
Disclosure of Invention
According to the problems existing in the prior art, the invention discloses a software vulnerability identification method with weight deviation based on graph neural network detection, which comprises the following steps:
s1, regarding words and symbols in source codes as a node, and representing the source codes by adopting a composition graph to obtain an initial characteristic value of each node and an initial characteristic of a connected edge of each graph;
s2, taking the generated composition diagram as input, wherein node characteristics in the diagram representation are initial states Z, adding Z-D0S after the node input characteristics when the node initial states Z are larger than the node input characteristic dimension D, transmitting and receiving information by edges of different nodes in the diagram representation, determining how much information of the previous moment and the current moment needs to be transmitted to the next moment by the information received by the nodes through the updating gate, determining how much information of the previous moment and the current moment needs to be discarded by the node received by the node through the resetting gate, finally combining the information output by the resetting gate, the information output by the updating gate and the information of the node, and outputting a node activation value by a Sigmoid activation function as the node state of the final moment;
s3, after the word nodes are fully updated, the words are aggregated to a graphic level representation of the function codes, and a final vulnerability recognition result is generated based on the representation;
s4, using
Figure BDA0003112340660000021
The coefficients minimize the defective software vulnerability loss value, thereby identifying the software vulnerability.
Further, the initial feature of the connected edge of each graph in S1 adopts a NL Graph Embedding method to obtain the edge weight, which is specifically as follows:
the method comprises the steps that a sliding window word co-occurrence mechanism is adopted to pattern a source code, edges in a formed graph are characteristic values, nodes appear in a sliding window, and the weight of the edges is increased by one;
all source codes are de-duplicated, each word is mapped to a node, and all sliding windows are obtained;
processing each sliding window, regarding each word in each sliding window as a node V, and counting the current sliding window W n One node V of (a) i With another node V j The number of times of co-occurrence between the two nodes, and adding the result to the previous number of times of co-occurrence of the two nodes as an edge initial feature E between the two nodes ij 。
Further, the following specific mode is adopted in S2:
for each node v in the graph i E V, initializing at the beginning, setting the initial state vector of each node as Z dimension, and inputting the characteristic x of each node i The dimension is D dimension, if the state vector dimension Z of the node is larger than the node input feature dimension D, Z-D0 s are added behind the node input feature;
setting T as the total number of time steps, wherein in each time step T is less than or equal to T, each node not only receives information of adjacent nodes, but also sends information to the adjacent nodes, and the adjacent relation between the nodes is represented by an adjacent matrix; the interaction between nodes is expressed as the following formula
Figure BDA0003112340660000031
Wherein A is v: Is two columns of in and out of the corresponding node selected from the matrix A, and the matrix A epsilon R D|v|×2D|v| ，A v: ∈R D|v|×2D ，/>
Figure BDA0003112340660000032
The D|v|dimension vector is formed by gathering the characteristics of all nodes at the time t-1, and b is deviation;
at time step t, a further calculation is made using the following formulaNew door
Figure BDA0003112340660000033
Wherein->
Figure BDA0003112340660000034
For the input vector of the t-th time step, a linear transformation is performed,/for the input vector of the t-th time step>
Figure BDA0003112340660000035
The information of the previous time step t-1 is stored, and the update gate adds the two parts of information and activates the two parts of information by using a Sigmoid activation function;
at time step t, we use
Figure BDA0003112340660000036
Equation calculation reset gate, wherein ∈>
Figure BDA0003112340660000037
Input vector for the t-th time step, < >>
Figure BDA0003112340660000038
The information of the previous time step t-1 is stored;
the new memory content will use the reset gate to store the past related information, and its calculation expression is as follows
Figure BDA0003112340660000039
Wherein->
Figure BDA00031123406600000310
For the input vector of the t-th time step, a linear transformation, right-hand matrix W, is applied by calculating the reset gate +.>
Figure BDA00031123406600000311
And->
Figure BDA00031123406600000312
Hadamard product of (A), i.e.)>
Figure BDA00031123406600000313
And->
Figure BDA00031123406600000314
Is obtained by means of a reset gate +.>
Figure BDA00031123406600000315
Controlling the opening size of the gate control, adding the current information and the information of the previous moment, and activating by using a hyperbolic tangent activation function tanh;
finally calculate
Figure BDA00031123406600000316
The vector will retain the information that the current cell passed to the next cell, its computational expression is +.>
Figure BDA00031123406600000317
Wherein->
Figure BDA00031123406600000318
And->
Figure BDA00031123406600000319
The Hadamard product of (2) represents the information that was retained until the last time step, +.>
Figure BDA00031123406600000320
And->
Figure BDA00031123406600000321
The Hadamard product of (2) represents the information that the current memory retains to the final memory, and the sum of the two parts of information is equal to the output content of the final gating loop unit.
Further, the following specific mode is adopted in S3: the calculation process of the graphical level representation of the function code is defined as follows:
Figure BDA0003112340660000041
Figure BDA0003112340660000042
wherein f 1 And f 2 Is two multi-layer perceptrons (MLP), f 1 As soft attention weight, f 2 As a nonlinear feature transformation, a graph representation h for each code sample G The application of a maximum pool function tag predicts +.>
Figure BDA0003112340660000043
Where W and b are weights and deviations.
By adopting the technical scheme, the software vulnerability recognition method with weight deviation based on the graph neural network detection can effectively represent the structural characteristics of the source code by converting the text format of the source code into the representation form of the source code graph, so that the model can learn the source code more fully. The number of defective samples in the source code is small, by
Figure BDA0003112340660000044
The coefficients replace the common cross entropy loss function to make the model pay more attention to defective samples, so that the quality of software is improved better.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of the embedding process of the present invention;
fig. 3 is a schematic diagram of the construction process of the adjacency matrix a in the present invention.
Detailed Description
In order to make the technical scheme and advantages of the present invention more clear, the technical scheme in the embodiment of the present invention is clearly and completely described below with reference to the accompanying drawings in the embodiment of the present invention:
the software vulnerability recognition method with weight deviation based on the graph neural network detection shown in fig. 1 specifically comprises the following steps:
step 1: graph embedding model
The word and symbol in the source code are regarded as one node, and the source code composition graph representation g= { V, E } is taken as the input of GGNN. V represents a set of points and E represents a set of edges. It is first necessary to obtain an initial feature value of each node, and also an initial feature of the connected edge of each graph. For the initial feature of the edge, the graph embedding model uses the method of NL Graph Embedding to obtain the edge weight. For the initial feature of the node, a Glove trained word embedding dictionary (glove.6b.300) is used to obtain an embedding vector for each word.
NL Graph Embedding: the source code is patterned using a sliding window word co-occurrence mechanism. As shown in fig. 2, a window size equal to 3 is used. For NL "static int has_duration (avformat context_n)", all words are first de-duplicated, each word is mapped to an index, and then all sliding windows are acquired. Thereafter, processing is performed for each sliding window. Each word in each sliding window is considered as a node V, and the current sliding window W is counted n One node V of (a) i With another node V j The number of co-occurrences between the two nodes is simply added to the previous number of co-occurrences of the two nodes as an edge initiation feature E between the two nodes ij 。
Step 2: gate control graph neural network prediction module
Graph g= { V, E } generated by graph embedding model is taken as input, for each node V in the graph i E V, initializing at the beginning, the formula is
Figure BDA0003112340660000051
Setting the initial state vector of each node to Z dimension, and inputting characteristic x of each node i The dimension is D dimension. If the state vector dimension Z of the node is greater than the node input feature dimension D, Z-D0 s are added behind the node input feature.
T is the total number of time steps, and in each time step T is less than or equal to T, each node can not only accept the information of the adjacent node, but also send the information to the adjacent node. The adjacency between nodes is represented by an adjacency matrix. The interaction between nodes is expressed as the following formula
Figure BDA0003112340660000052
Wherein A is v: Is two columns in and out of the corresponding node selected from the matrix A of FIG. 3 (c), matrix A εR D|v|×2D|v| ，A v: ∈R D|v|×2D 。/>
Figure BDA0003112340660000053
Is a vector of dimension d|v| formed by aggregating the features of all nodes at time t-1. In addition, b is the deviation.
At time step t, we first calculate the update gate using the following formula
Figure BDA0003112340660000054
Wherein->
Figure BDA0003112340660000055
For the input vector of the t-th time step, a linear transformation (with weight matrix W z Multiplication). />
Figure BDA0003112340660000056
The information of the previous time step t-1 is stored and is also subjected to a linear transformation. The update gate adds the two pieces of information and activates using the Sigmoid activation function. The update gate can help the model decide how much past information to pass to the future, or how much information from the previous and current time steps needs to continue to pass. In this way the model can decide to replicate all the information from the past and thus reduce the risk of gradient extinction.
At time step t, we calculate the reset gate using the following formula
Figure BDA0003112340660000061
Wherein->
Figure BDA0003112340660000062
Input vector for the t-th time step, < >>
Figure BDA0003112340660000063
The information of the previous time step t-1 is stored. Resetting the gate is similar to updating, but the parameters and usefulness of the linear transformation are different, and he can help the model decide how much past information needs to be forgotten.
The new memory content will use the reset gate to store the past related information, its calculation expression is
Figure BDA0003112340660000064
Wherein->
Figure BDA0003112340660000065
The input vector for the t-th time step is subjected to a linear transformation, i.e. a right-hand multiplication matrix W. By calculating reset gate +.>
Figure BDA0003112340660000066
And->
Figure BDA0003112340660000067
Hadamard product of (A), i.e.)>
Figure BDA0003112340660000068
And->
Figure BDA0003112340660000069
And the corresponding element product of (c) to obtain the data after "reset". By resetting the gate->
Figure BDA00031123406600000610
The magnitude of the gating opening is controlled so the Hadamard product will determine the previous information to be retained and forgotten. The two parts are then combinedAfter addition of the calculation results of (2), activation is performed using a hyperbolic tangent activation function tanh.
At the end of the calculation
Figure BDA00031123406600000611
The vector will retain the information that the current cell passed to the next cell, its computational expression is +.>
Figure BDA00031123406600000612
Wherein->
Figure BDA00031123406600000613
And->
Figure BDA00031123406600000614
The Hadamard product of (2) represents the information that was retained until the last time step, +.>
Figure BDA00031123406600000615
And->
Figure BDA00031123406600000616
The Hadamard product of (c) represents the information that the current memory retains to the final memory. The two pieces of information add up to the content output by the final gating loop.
Step 3 readout Module
After the word nodes are sufficiently updated, they are aggregated into a graphical level representation of the function code and a final prediction is generated based on the representation. The calculation process of the graphical level representation of the function code is defined as follows:
Figure BDA00031123406600000617
Figure BDA00031123406600000618
wherein f 1 And f 2 Is two multi-layer perceptron (MLP). The former acts as a soft attention weight and the latter acts as a nonlinear feature transformation. In addition to the average weighted word features, we also represent h for each code sample graph G Applying a maximum ofChi Hanshu, the idea is that each word plays a role in the code and the contribution of the keywords should be more explicit. Finally, the tag predicts +_by inputting the level vector into the softmax layer>
Figure BDA0003112340660000071
Where W and b are weights and deviations.
Step 4 weighting model
In practical application of code defect prediction, defective samples often occupy a few of all samples, and some defective code samples are misreported and not reported, because previous work usually uses a cross entropy loss function to measure model performance, the ratio of the loss function weight of the defective samples to the non-defective samples is predicted to be 1:1, which results in insufficient learning of defective sample characteristics with small occupied ratio and difficult recognition as full as possible. We consider the cross entropy loss function to be responsible for this problem. Let y denote the actual label as such,
Figure BDA0003112340660000072
representing the predicted output, the original cross entropy loss function is +.>
Figure BDA0003112340660000073
Therefore, we use +.>
Figure BDA0003112340660000074
Coefficients replace the cross entropy loss function. At->
Figure BDA0003112340660000075
In the coefficients>
Figure BDA0003112340660000076
Acting as a scaling factor, for simple samples (tending to 1 or 0) the model is made to pay less attention to them. Indeed, from the derivative point of view, once the model correctly classifies the current sample (just passing 0.5), DSC will make the model pay less attention to it, rather than encourage the model to approach 0 or like the cross entropy loss function1, which effectively avoids model training from being dominated by simple samples due to excessive simple samples. />
Figure BDA0003112340660000077
Coefficients of
Figure BDA0003112340660000078
The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.

Claims (1)

1. A software vulnerability recognition method based on a graph neural network detection with weight deviation is characterized by comprising the following steps:
s1, regarding words and symbols in source codes as a node, and representing the source codes by adopting a composition graph to obtain an initial characteristic value of each node and an initial characteristic of a connected edge of each graph;
s2, taking the generated composition diagram as input, wherein node characteristics in the diagram representation are initial states Z, adding Z-D0S after the node input characteristics when the node initial states Z are larger than the node input characteristic dimension D, transmitting and receiving information by edges of different nodes in the diagram representation, determining how much information of the previous moment and the current moment needs to be transmitted to the next moment by the information received by the nodes through the updating gate, determining how much information of the previous moment and the current moment needs to be discarded by the node received by the node through the resetting gate, finally combining the information output by the resetting gate, the information output by the updating gate and the information of the node, and outputting a node activation value by a Sigmoid activation function as the node state of the final moment;
s3, after the word nodes are fully updated, the words are aggregated to a graphic level representation of the function codes, and a final vulnerability recognition result is generated based on the representation;
s4, using
Figure FDA0004238484440000011
-the Dice coefficient minimizes a defective software vulnerability loss value, thereby identifying a software vulnerability;
the initial characteristics of the connected edges of each graph in S1 are obtained by adopting an NLGraphEmbedding method, and the specific mode is as follows:
the method comprises the steps that a sliding window word co-occurrence mechanism is adopted to pattern a source code, edges in a formed graph are characteristic values, nodes appear in a sliding window, and the weight of the edges is increased by one;
all source codes are de-duplicated, each word is mapped to a node, and all sliding windows are obtained;
processing each sliding window, regarding each word in each sliding window as a node V, and counting the current sliding window W n One node V of (a) i With another node V j The number of times of co-occurrence between the two nodes, and adding the result to the previous number of times of co-occurrence of the two nodes as an edge initial feature E between the two nodes ij
S2 specifically adopts the following mode:
for each node v in the graph i E V, initializing at the beginning, setting the initial state vector of each node as Z dimension, and inputting the characteristic x of each node i The dimension is D dimension, if the state vector dimension Z of the node is larger than the node input feature dimension D, Z-D0 s are added behind the node input feature;
setting T as the total number of time steps, wherein in each time step T is less than or equal to T, each node not only receives information of adjacent nodes, but also sends information to the adjacent nodes, and the adjacent relation between the nodes is represented by an adjacent matrix; the interaction between nodes is expressed as the following formula
Figure FDA0004238484440000021
Wherein A is v: Is two columns of in and out of the corresponding node selected from the matrix A, and the matrix A epsilon R D|v|×2D|v| ，A v: ∈R D|v|×2D ，/>
Figure FDA0004238484440000022
The D|v|dimension vector is formed by gathering the characteristics of all nodes at the time t-1, and b is deviation;
at time step t, an update gate is calculated using the following formula
Figure FDA0004238484440000023
Wherein->
Figure FDA0004238484440000024
For the input vector of the t-th time step, a linear transformation is performed,/for the input vector of the t-th time step>
Figure FDA0004238484440000025
The information of the previous time step t-1 is stored, and the update gate adds the two parts of information and activates the two parts of information by using a Sigmoid activation function;
at time step t, we use
Figure FDA0004238484440000026
Equation calculation reset gate, wherein ∈>
Figure FDA0004238484440000027
Input vector for the t-th time step, < >>
Figure FDA0004238484440000028
The information of the previous time step t-1 is stored;
the new memory content will use the reset gate to store the past related information, and its calculation expression is as follows
Figure FDA0004238484440000029
Wherein->
Figure FDA00042384844400000210
For the input vector of the t-th time step, a linear transformation, right-hand matrix W, is applied by calculating the reset gate +.>
Figure FDA00042384844400000211
And->
Figure FDA00042384844400000212
Hadamard product of (A), i.e.)>
Figure FDA00042384844400000213
And->
Figure FDA00042384844400000214
Is obtained by means of a reset gate +.>
Figure FDA00042384844400000215
Controlling the opening size of the gate control, adding the current information and the information of the previous moment, and activating by using a hyperbolic tangent activation function tanh;
finally calculate
Figure FDA00042384844400000216
The vector will retain the information that the current cell passed to the next cell, its computational expression is
Figure FDA00042384844400000217
Wherein->
Figure FDA00042384844400000218
And->
Figure FDA00042384844400000219
The Hadamard product of (2) represents the information that was retained until the last time step, +.>
Figure FDA00042384844400000220
And->
Figure FDA00042384844400000221
The Hadamard product of (2) represents the information which is reserved from the current memory to the final memory, and the sum of the two parts of information is equal to the output content of the final gating circulation unit;
s3, specifically adopting the following modes: the calculation process of the graphical level representation of the function code is defined as follows:
Figure FDA00042384844400000222
wherein f 1 And f 2 Is two multi-layer perceptrons (MLP), f 1 As soft attention weight, f 2 As a nonlinear feature transformation, a graph representation h for each code sample G The application of a maximum pool function tag predicts +.>
Figure FDA0004238484440000031
Where W and b are weights and deviations.
CN202110652749.4A 2021-06-11 2021-06-11 Software vulnerability identification method based on graph neural network detection with weight deviation Active CN113378176B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110652749.4A CN113378176B (en) 2021-06-11 2021-06-11 Software vulnerability identification method based on graph neural network detection with weight deviation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110652749.4A CN113378176B (en) 2021-06-11 2021-06-11 Software vulnerability identification method based on graph neural network detection with weight deviation

Publications (2)

Publication Number Publication Date
CN113378176A CN113378176A (en) 2021-09-10
CN113378176B true CN113378176B (en) 2023-06-23

Family

ID=77574117

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110652749.4A Active CN113378176B (en) 2021-06-11 2021-06-11 Software vulnerability identification method based on graph neural network detection with weight deviation

Country Status (1)

Country Link
CN (1) CN113378176B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111259394A (en) * 2020-01-15 2020-06-09 中山大学 Fine-grained source code vulnerability detection method based on graph neural network
CN111274134A (en) * 2020-01-17 2020-06-12 扬州大学 Vulnerability identification and prediction method and system based on graph neural network, computer equipment and storage medium
CN112560049A (en) * 2020-12-28 2021-03-26 苏州极光无限信息技术有限公司 Vulnerability detection method and device and storage medium

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018184102A1 (en) * 2017-04-03 2018-10-11 Royal Bank Of Canada Systems and methods for malicious code detection
US11568055B2 (en) * 2019-08-23 2023-01-31 Praetorian System and method for automatically detecting a security vulnerability in a source code using a machine learning model

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111259394A (en) * 2020-01-15 2020-06-09 中山大学 Fine-grained source code vulnerability detection method based on graph neural network
CN111274134A (en) * 2020-01-17 2020-06-12 扬州大学 Vulnerability identification and prediction method and system based on graph neural network, computer equipment and storage medium
CN112560049A (en) * 2020-12-28 2021-03-26 苏州极光无限信息技术有限公司 Vulnerability detection method and device and storage medium

Also Published As

Publication number Publication date
CN113378176A (en) 2021-09-10

Similar Documents

Publication Publication Date Title
CN111897908B (en) Event extraction method and system integrating dependency information and pre-training language model
Sharma et al. Efficient Classification for Neural Machines Interpretations based on Mathematical models
CN111046668B (en) Named entity identification method and device for multi-mode cultural relic data
CN107506414A (en) A kind of code based on shot and long term memory network recommends method
CN111985245A (en) Attention cycle gating graph convolution network-based relation extraction method and system
US20210081841A1 (en) Visually creating and monitoring machine learning models
CN112364352B (en) Method and system for detecting and recommending interpretable software loopholes
CN112883714B (en) ABSC task syntactic constraint method based on dependency graph convolution and transfer learning
CN117236677A (en) RPA process mining method and device based on event extraction
CN118276913B (en) Code completion method based on artificial intelligence
CN113742733A (en) Reading understanding vulnerability event trigger word extraction and vulnerability type identification method and device
CN113378178B (en) Deep learning-based graph self-confidence learning software vulnerability detection method
CN117591913A (en) Statement level software defect prediction method based on improved R-transducer
CN113378176B (en) Software vulnerability identification method based on graph neural network detection with weight deviation
CN116151375B (en) Event tracing reasoning method based on inverse facts and path mining
CN116932382A (en) Software defect prediction method based on position sensitive hierarchical attention network
CN116069337A (en) Code defect automatic repair method combining repair template and deep learning
CN113051607B (en) Privacy policy information extraction method
CN113867724A (en) Method and system for automatically generating GUI (graphical user interface) code, server and medium
CN111562943A (en) Code clone detection method and device based on event embedded tree and GAT network
Zhao et al. Neural-Network-Based Feature Learning: Recurrent Neural Network
CN113377422B (en) Self-recognition technical liability method based on deep learning identification
CN118672594B (en) Software defect prediction method and system
CN113901215B (en) Text inclusion recognition method fusing high-level semantic information and low-level semantic information
CN114090724A (en) User data processing method, system, equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant