CN108549817A - A kind of software security flaw prediction technique based on text deep learning - Google Patents
A kind of software security flaw prediction technique based on text deep learning Download PDFInfo
- Publication number
- CN108549817A CN108549817A CN201810353774.0A CN201810353774A CN108549817A CN 108549817 A CN108549817 A CN 108549817A CN 201810353774 A CN201810353774 A CN 201810353774A CN 108549817 A CN108549817 A CN 108549817A
- Authority
- CN
- China
- Prior art keywords
- software
- text
- word
- loophole
- source code
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 36
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 27
- 238000010801 machine learning Methods 0.000 claims abstract description 20
- 238000003062 neural network model Methods 0.000 claims abstract description 8
- 239000013598 vector Substances 0.000 claims description 14
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 9
- 239000000284 extract Substances 0.000 claims description 2
- 101100124609 Caenorhabditis elegans zyg-12 gene Proteins 0.000 claims 2
- 238000005259 measurement Methods 0.000 description 18
- 238000012549 training Methods 0.000 description 15
- 230000008569 process Effects 0.000 description 10
- 238000012360 testing method Methods 0.000 description 9
- 210000002569 neuron Anatomy 0.000 description 7
- 238000007477 logistic regression Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 230000009467 reduction Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000000843 powder Substances 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2221/00—Indexing scheme relating to security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F2221/03—Indexing scheme relating to G06F21/50, monitoring users, programs or devices to maintain the integrity of platforms
- G06F2221/033—Test or assess software
Abstract
The present invention is a kind of software security flaw prediction technique based on text deep learning, using deep neural network model and shallow-layer machine learning algorithm from history software source code text learning characteristic and knowledge, can be used in predicting the security breaches in new software source code.The present invention uses the Structural Characteristics in deep neural network model learning software source code text feature, using the feature learnt as the input of grader, adjustment is trained to grader, obtains optimal loophole prediction model, the loophole prediction of the new software module for the software.
Description
Technical field
The present invention relates to a kind of software security flaw prediction techniques based on text deep learning, belong to software security flaw
Electric powder prediction.
Background technology
Software security flaw prediction can know in advance in software source code module there is a possibility that software vulnerability or
Quantity, for software developer according to prediction result, limited time and funds can targetedly put into those there are loopholes can
Can property it is high and quantity more than software module test in, the efficiency of software test is improved with this.
Currently, common software vulnerability prediction establishes software vulnerability prediction model using shallow-layer machine learning method, establish
Process is as shown in Figure 1：
1. the measurement metric of software source code module is established
Currently, the measurement metric of software source code module is established, there are mainly two types of methods：A kind of method is to use metric software
The index (metrics) of code quality be used as measurement metric, such as by the CK indexs of object-oriented program (WMC, DIT, NOC, CBO,
RFC and LCOM etc.), the code revision characteristic index of software-oriented development process, developer's experience level index, intermodule
Dependency degree index and project team organizational structure reasonability index etc. are used as measurement metric, we are referred to as based on software index
Measurement metric；Another method is that software source code is considered as text, and the frequency that each word occurs in text is considered as measurement metric,
We are referred to as the measurement metric based on code text word.
2. software history vulnerability scan is established
It is collected from the software vulnerability library having disclosed and is directed to the leaky to current institute of some software project, establish needle
To the software vulnerability library of the software project.It is specified in software vulnerability library for loophole in each software module of the software project
Position and quantity.
Software vulnerability library provides the historical knowledge of the loophole distribution for the software project.
3. software vulnerability predicts the training of machine learning model, test
For some software project, the concrete numerical value of the measurement metric index of each software module is calculated, software history is utilized
Vulnerability database obtains the label or defects count of the presence or absence of each software module loophole, and then selection is suitable for the software project
(there is presently no disclosed data to show to have used deep learning algorithm, and what is used is all based on shallow-layer for machine learning algorithm
The algorithm of habit), using the concrete numerical value of measurement metric index as input, whether there is or not the labels in gaps and omissions hole or loophole quantity as defeated
Go out, establishes the machine learning model that (training, test and parameter adjustment) is directed to the software vulnerability prediction of the software project.
4. software vulnerability predicts the application of machine learning model
Training is completed and the loophole prediction model of test passes can be to the loophole feelings of the new software module of the project
Condition is predicted.The concrete numerical value for calculating the measurement metric index of new software module first, numerical value is input in prediction model,
The result of model calculation output is that the software module there is a possibility that loophole or the quantity there are loophole.
In entire modeling process, aspect there are three the factors of prediction model performance is influenced：Measurement metric index is selected
It takes, the quality of vulnerability database and specific machine learning algorithm.Selected measurement metric index should can reflect leaky mould
Block and substantive characteristics without loophole module, that is, measurement metric index is to leaky and without loophole module have certain area
The ability of dividing；The quality in software vulnerability library also largely influences the performance of model, and the software vulnerability library established should
With higher accuracy and preferable completeness；Machine learning algorithm itself also has different performances, for different soft
Part project, selection are suitable for this item purpose machine learning algorithm.Factor in terms of these three be connect each other it is interactional, it is comprehensive
Close the performance for determining this prediction model.
In the case where the quality of history vulnerability database is set, for certain class measurement metric index, different machine learning algorithms
Performance difference it is larger, carried out a large amount of trial for different machine learning algorithms at present, but the algorithm of all uses
Belong to shallow-layer machine learning method, does not there is disclosed data to show that someone uses deep learning method.
Existing data shows that in most cases the prediction effect of the measurement metric based on code text word wants bright
The aobvious prediction effect better than the measurement metric based on software index.Actual conditions are found, measurement is being used as using code text word
In the prediction technique of member, the dimension of measurement metric would generally be very high, and sometimes even over 10,000 dimensions, and these dimensional characteristics are sparse
's.In such cases, at present frequently with based on shallow-layer study machine learning algorithm, for deep learning, feature
Learning effect is poor, directly results in the effect Shortcomings of loophole prediction.A large amount of research and practice have shown that, characteristic dimension very
In the case of high and sparsity is very strong, the learning ability of deep learning will be substantially better than shallow-layer study, this is because deep learning
It can learn the Structural Characteristics in high dimensional feature automatically, achieve the purpose that carry out compression and dimensionality reduction to feature.
Invention content
The present invention is a kind of software security flaw prediction technique based on text deep learning, using deep neural network mould
Type and shallow-layer machine learning algorithm learning characteristic and knowledge from history software source code text, can be used in new software source
Security breaches in code are predicted.
The invention is realized by the following technical scheme：
A kind of software security flaw prediction technique based on text deep learning, for the source code of software, with therein
Software module is processing unit, counts the number that each word occurs in entire text, and occurrence number is normalized
Processing obtains the frequency of word appearance, and the frequency that word and word are occurred is as the feature vector of the source code text, the spy
Input of the sign vector as the feature learning device based on deep neural network structure, feature learning device learn it to obtain spy
The Structural Characteristics of vector are levied, input of the Structural Characteristics as grader is trained adjustment to the parameter of grader, obtains
Optimal loophole prediction model is obtained, the loophole prediction of the new software module for the software.
Further, described eigenvector is extracted in the following ways：To predict that the software module in object is that processing is single
Member is first rejected the punctuation mark and code annotation that occur in source code text, is made with space to remaining text
Each word is extracted for separator, and counts the number that each word occurs in entire text, will finally go out occurrence
Number is normalized, and the result of processing is the frequency that word occurs, and thus obtains the feature for the source code text
Vector.
Further, the feature vector of source code text uses following representation：
ComponentName：(Item_1：Number_1；Item_2：Number_2；…；Item_n：
Number_n)
Wherein, ComponentName indicates that the title of the module, Item_1 indicate the title of the 1st word,
Number_1 indicates that the frequency that the 1st word occurs, Item_2 indicate that the title of the 2nd word, Number_2 indicate the 2nd
The frequency that a word occurs, this is analogized, and Item_n indicates that the title of n-th of word, Number_n indicate that n-th of word goes out
Existing frequency.
Further, the grader uses shallow-layer machine learning algorithm.
Further, the output of the grader is two classification, that is, whether the module of predicted object has loophole.
Further, it is established using different deep neural network models and different classifier algorithms for prediction object
A variety of loophole prediction models, contrast properties index value determine optimal loophole prediction model.
Beneficial effects of the present invention：The present invention uses in deep neural network model learning software source code text feature
Structural Characteristics, then using the feature learnt as the input of grader, its advantage is that software source code textual words feature to
Amount is that a kind of dimension is very high and the very strong feature of sparsity, and existing shallow-layer machine learning algorithm processing capacity is limited, using depth
Degree learning algorithm learns to obtain Structural Characteristics first, achievees the purpose that Feature Dimension Reduction, then the feature that study is obtained as shallow
The input of layer learning algorithm, can improve the performance indicator of loophole prediction model in this way so that loophole prediction model has precision
The relatively low advantage of height, false alarm rate and false dismissed rate.
Description of the drawings
Fig. 1 is that software vulnerability prediction uses shallow-layer machine learning method flow chart in the prior art；
Fig. 2 is the software security flaw prediction technique flow chart based on text deep learning of the present invention；
Fig. 3 is the structure chart of loophole prediction model in the present invention；
Fig. 4 is the feature learning device based on stack self-encoding encoder structure in the present invention.
Specific implementation mode
The invention will be described further below in conjunction with the accompanying drawings.
As shown in Fig. 2, the software security flaw prediction technique based on text deep learning of the present invention includes the following steps：
Step 1: software source code textual words feature extraction
For the source code (predicting object) of some software project, using software module therein as processing unit, first
The punctuation mark and code annotation that occur in module source code text are rejected, to remaining text using space as point
Each word extracted every symbol, and counts the number that each word occurs in entire text, finally by occurrence number into
Row normalized, the result of processing are the frequency that word occurs, and thus obtain the feature for the module source code text
Vector is set as following representation：
ComponentName：(Item_1：Number_1；Item_2：Number_2；…；Item_n：
Number_n)
Wherein, ComponentName indicates that the title of the module, Item_1 indicate the title of the 1st word,
Number_1 indicates that the frequency that the 1st word occurs, Item_2 indicate that the title of the 2nd word, Number_2 indicate the 2nd
The frequency that a word occurs, this is analogized, and Item_n indicates that the title of n-th of word, Number_n indicate that n-th of word goes out
Existing frequency.
Step 2: establishing software history vulnerability database
All disclosed software vulnerabilities for the software project (prediction object) are collected from disclosed software vulnerability library,
These disclosed loophole data give the specific location in the software module residing for each loophole simultaneously, thus can determine each
The history loophole quantity that software module occurred, referred to as the loophole label of the software module.If being directed to some software mould
Block does not find the open loophole for having for the module, then it is assumed that the loophole quantity of the software module is zero.For loophole label
It is further processed：If loophole label is not equal to 0, it is uniformly set as 1, if loophole label is 0, remains set to 0, this
Sample, otherwise loophole label is 1 or be 0, only indicate whether there is or not.
Step 3: establishing the feature learning device based on deep neural network structure
On the basis of step 1, selection is suitable for the deep neural network model of textual words feature learning.Software source
The key property of code text word is between word there is no semantic relation, and the relevance between these features is not strong, because
There are many this suitable deep neural network models, for one such, such as with the depth of stack self-encoding encoder structure
For neural network, structure is as shown in Figure 4, wherein input layer has n input feature vector amount x1, x2..., xn, corresponding respectively to walk
Word feature the amount Item_1, Item_2 ..., Item_n extracted in rapid one, after the quantity n of word feature amount is determined, choosing
Corresponding neural network input terminal is selected, such as word feature amount there are 5, then selects corresponding 5 of input terminal as input, hidden layer
With h layers, the quantity of every layer of neuron is k respectively1、k2、…、kh, the quantity of hidden neuron successively reduces, i.e. k1＜ k2
＜ ... ＜ kh.The quantity of the neuron of output layer is m.It is to connect entirely between every layer of neuron, the neuron of input layer is not joined
With calculating.The activation primitive of neuron uses Sigmoid functions.Neuron in the number of plies of hidden layer and every layer in feature learning device
Quantity specifically determined according to actual conditions.
Step 4: establishing grader
On the basis of step 3, the grader of loophole prediction model is established, the input of grader is established in step 3
Feature learning device output, that is, the feature that arrives of third step middle school's acquistion.Grader uses shallow-layer machine learning algorithm, can
There are many shallow-layer machine learning algorithms of selection, such as LogisticRegression algorithms, SVM, NN, Bayesian network etc.；Point
The output of class device is classified for two, i.e., and 0 or 1, indicate whether the module of predicted object has loophole, when output is not 0 or 1,
The then parameter of neural network model or the model parameter of grader in feature learning device in set-up procedure three, until output for 0 or
1.It was proved that the model parameter for adjusting grader is easier optimization to obtain output to be 0 or 1.
Step 5: being trained and testing to prediction model
Step 1: two, three, four on the basis of is trained and tests to prediction model, the specific method is as follows：Utilize step
Rapid one method extracts the Text eigenvector of each software module, and the history that prediction object is established using the method for step 2 is leaked
Cave depot, to obtain prediction object each module loophole label.It is soft that all feature vectors and corresponding label constitute this
The sample database of part project.Sample database is divided into two parts, a part is training sample for training；A part is used to test, and is
Test sample.Using the Text eigenvector of training sample as the input of step 3, feature learning device that step 3 is established
Every layer is individually trained.Training method uses the training algorithm corresponding to used deep neural network structure.This process
Referred to as pre-training, pre-training process are both the learning process of the dimensionality reduction and Structural Characteristics to primitive character.It is self-editing with stack
For the deep neural network of code device structure, successively it is trained based on BP algorithm using greedy method, training process is fallen into a trap
It is equal with input value to calculate the output valve used when reverse propagated error.
After training process, the input by the output of feature learning device as the grader in step 4, to grader
Parameter be trained adjustment, according to the criterion corresponding to specific learning algorithm, obtain the parameter value of optimal grader.
By taking LogisticRegression graders as an example, using the output of feature learning device as LogisticRegression graders
Input, LogisticRegression classifier parameters are adjusted using gradient descent algorithm, when cost function obtain
Both the optimized parameter of LogisticRegression graders is obtained when minimum value, the training process of grader terminates at this time.
After pre-training and classifier parameters adjustment process, the training of entire loophole prediction model terminates.Next
The performance of model is tested using test sample, obtains the occurrence of performance indicator, these performance indicators include precision, standard
Exactness, false alarm rate, false dismissed rate etc..
Step 6: determining optimum prediction model according to performance indicator
For prediction object, established using different deep neural network structural models and different classifier algorithms a variety of
Loophole prediction model obtains the performance index value of each prediction model on the basis of step 5.Contrast properties index, according to
Actual demand determines the optimal loophole prediction model for being suitable for this software project.For example, in certain situations it is desirable to by smart
Degree index is placed above the other things, at this time the highest loophole prediction model of choice accuracy index；And in some cases, need by
Accuracy index is placed above the other things, answers the highest loophole prediction model of accuracy of selection index at this time.
Step 7: prediction model is applied to new software module
The prediction model of the best performance obtained in step 6 is applied to the leakage of the new software module of the software project
It predicts in hole.New software module feature vector value is calculated first, using result of calculation as the input of prediction model, model running
Output is whether the new software module has leaky label.
Claims (6)
1. a kind of software security flaw prediction technique based on text deep learning, which is characterized in that it is directed to the source code of software,
Using software module therein as processing unit, each word frequency of occurrences in source code text is counted, word and word are gone out
Feature vector of the existing frequency as the source code text, this feature vector is as the characterology based on deep neural network structure
The input of device is practised, feature learning device learns it to obtain the Structural Characteristics of feature vector, which, which is used as, divides
The input of class device is trained adjustment to the parameter of grader, obtains optimal loophole prediction model, for the new of the software
The loophole of software module is predicted.
2. a kind of software security flaw prediction technique based on text deep learning as described in claim 1, which is characterized in that
Described eigenvector is extracted in the following ways：To predict the software module in object as processing unit, first to source code text
The punctuation mark and code annotation occurred in this is rejected, to remaining text using space as separator by each word
It extracts, and counts the number that each word occurs in entire text, finally occurrence number is normalized, locate
The result of reason is the frequency that word occurs, and thus obtains the feature vector for the source code text.
3. a kind of software security flaw prediction technique based on text deep learning as claimed in claim 1 or 2, feature exist
In the feature vector of source code text uses following representation：
Com ponentN am e：(Item_1：N um ber_1；Item_2：N um ber_2；；…；Item_n：N
Um ber_n)
Wherein, Com ponentN ame indicate that the title of the module, Item_1 indicate the title of the 1st word, N um
Ber_1 indicates that the frequency that the 1st word occurs, Item_2 indicate that the title of the 2nd word, N um ber_2 indicate the 2nd
The frequency that word occurs, this is analogized, and Item_n indicates that the title of n-th of word, N um ber_n indicate that n-th of word goes out
Existing frequency.
4. a kind of software security flaw prediction technique based on text deep learning as claimed in claim 1 or 2, feature exist
In the grader uses shallow-layer machine learning algorithm.
5. a kind of software security flaw prediction technique based on text deep learning as claimed in claim 4, which is characterized in that
The output of the grader is two classification, that is, whether the module of predicted object has loophole.
6. a kind of software security flaw prediction technique based on text deep learning as claimed in claim 1 or 2, feature exist
In this method further comprises, different deep neural network models and different classifier algorithms are used for prediction object
A variety of loophole prediction models are established, contrast properties index value determines optimal loophole prediction model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810353774.0A CN108549817A (en) | 2018-04-19 | 2018-04-19 | A kind of software security flaw prediction technique based on text deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810353774.0A CN108549817A (en) | 2018-04-19 | 2018-04-19 | A kind of software security flaw prediction technique based on text deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108549817A true CN108549817A (en) | 2018-09-18 |
Family
ID=63515568
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810353774.0A Pending CN108549817A (en) | 2018-04-19 | 2018-04-19 | A kind of software security flaw prediction technique based on text deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108549817A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109698823A (en) * | 2018-11-29 | 2019-04-30 | 广东电网有限责任公司信息中心 | A kind of Cyberthreat discovery method |
| CN109871688A (en) * | 2018-09-21 | 2019-06-11 | 中国人民解放军国防科技大学 | vulnerability threat degree evaluation method |
| CN111338692A (en) * | 2018-12-18 | 2020-06-26 | 北京奇虎科技有限公司 | Vulnerability classification method and device based on vulnerability codes and electronic equipment |
| CN111597301A (en) * | 2020-04-24 | 2020-08-28 | 北京百度网讯科技有限公司 | Text prediction method and device and electronic equipment |
| CN111897946A (en) * | 2020-07-08 | 2020-11-06 | 扬州大学 | Vulnerability patch recommendation method, system, computer equipment and storage medium |
| CN113239360A (en) * | 2021-04-30 | 2021-08-10 | 杭州安恒信息技术股份有限公司 | Network asset management method based on machine learning and related components |
| CN113779590A (en) * | 2021-09-16 | 2021-12-10 | 中国民航大学 | Source code vulnerability detection method based on multi-dimensional representation |
| US11429352B2 (en) | 2020-07-01 | 2022-08-30 | International Business Machines Corporation | Building pre-trained contextual embeddings for programming languages using specialized vocabulary |
| CN111338692B (en) * | 2018-12-18 | 2024-04-16 | 北京奇虎科技有限公司 | Vulnerability classification method and device based on vulnerability codes and electronic equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273752A (en) * | 2017-06-26 | 2017-10-20 | 北京理工大学 | Leak automatic classification method based on word frequency statisticses and naive Bayesian Fusion Model |
| CN107885999A (en) * | 2017-11-08 | 2018-04-06 | 华中科技大学 | A kind of leak detection method and system based on deep learning |
-
2018
- 2018-04-19 CN CN201810353774.0A patent/CN108549817A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273752A (en) * | 2017-06-26 | 2017-10-20 | 北京理工大学 | Leak automatic classification method based on word frequency statisticses and naive Bayesian Fusion Model |
| CN107885999A (en) * | 2017-11-08 | 2018-04-06 | 华中科技大学 | A kind of leak detection method and system based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| 董枫等: "面向Android二进制代码的缺陷预测方法", 《北京邮电大学学报》 * |
| 郑铁男等: "《数字编辑运营实训教程》", 30 September 2017, 北京：知识产权出版社 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109871688A (en) * | 2018-09-21 | 2019-06-11 | 中国人民解放军国防科技大学 | vulnerability threat degree evaluation method |
| CN109698823A (en) * | 2018-11-29 | 2019-04-30 | 广东电网有限责任公司信息中心 | A kind of Cyberthreat discovery method |
| CN111338692A (en) * | 2018-12-18 | 2020-06-26 | 北京奇虎科技有限公司 | Vulnerability classification method and device based on vulnerability codes and electronic equipment |
| CN111338692B (en) * | 2018-12-18 | 2024-04-16 | 北京奇虎科技有限公司 | Vulnerability classification method and device based on vulnerability codes and electronic equipment |
| CN111597301A (en) * | 2020-04-24 | 2020-08-28 | 北京百度网讯科技有限公司 | Text prediction method and device and electronic equipment |
| CN111597301B (en) * | 2020-04-24 | 2023-06-09 | 北京百度网讯科技有限公司 | Text prediction method and device and electronic equipment |
| US11429352B2 (en) | 2020-07-01 | 2022-08-30 | International Business Machines Corporation | Building pre-trained contextual embeddings for programming languages using specialized vocabulary |
| CN111897946A (en) * | 2020-07-08 | 2020-11-06 | 扬州大学 | Vulnerability patch recommendation method, system, computer equipment and storage medium |
| CN111897946B (en) * | 2020-07-08 | 2023-09-19 | 扬州大学 | Vulnerability patch recommendation method, vulnerability patch recommendation system, computer equipment and storage medium |
| CN113239360A (en) * | 2021-04-30 | 2021-08-10 | 杭州安恒信息技术股份有限公司 | Network asset management method based on machine learning and related components |
| CN113779590A (en) * | 2021-09-16 | 2021-12-10 | 中国民航大学 | Source code vulnerability detection method based on multi-dimensional representation |
| CN113779590B (en) * | 2021-09-16 | 2023-09-08 | 中国民航大学 | Source code vulnerability detection method based on multidimensional characterization |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108549817A (en) | A kind of software security flaw prediction technique based on text deep learning | |
| CN109408389B (en) | Code defect detection method and device based on deep learning | |
| CN109615116A (en) | A kind of telecommunication fraud event detecting method and detection system | |
| CN110309514A (en) | A kind of method for recognizing semantics and device | |
| CN113407694B (en) | Method, device and related equipment for detecting ambiguity of customer service robot knowledge base | |
| WO2021139279A1 (en) | Data processing method and apparatus based on classification model, and electronic device and medium | |
| CN108051660A (en) | A kind of transformer fault combined diagnosis method for establishing model and diagnostic method | |
| CN106528417A (en) | Intelligent detection method and system of software defects | |
| CN113672931B (en) | Software vulnerability automatic detection method and device based on pre-training | |
| CN113516228B (en) | Network anomaly detection method based on deep neural network | |
| CN109034140A (en) | Industrial control network abnormal signal detection method based on deep learning structure | |
| CN112308316A (en) | Crime number prediction method based on linear regression algorithm | |
| CN111950645A (en) | Method for improving class imbalance classification performance by improving random forest | |
| CN115687925A (en) | Fault type identification method and device for unbalanced sample | |
| CN111160959A (en) | User click conversion estimation method and device | |
| CN111046930A (en) | Power supply service satisfaction influence factor identification method based on decision tree algorithm | |
| CN109462578A (en) | Threat intelligence use and propagation method based on statistical learning | |
| CN107016416B (en) | Data classification prediction method based on neighborhood rough set and PCA fusion | |
| CN115145817A (en) | Software testing method, device, equipment and readable storage medium | |
| CN113343123B (en) | Training method and detection method for generating confrontation multiple relation graph network | |
| KR101703972B1 (en) | System and method for predicting groundwater potential area using spatial information | |
| CN112966778A (en) | Data processing method and device for unbalanced sample data | |
| CN117093922A (en) | Improved SVM-based complex fluid identification method for unbalanced sample oil reservoir | |
| CN116975881A (en) | LLVM (LLVM) -based vulnerability fine-granularity positioning method | |
| US20230252282A1 (en) | Method, server, and system for deep metric learning per hierarchical steps of multi-labels and few-shot inference using the same |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180918 |
|
| RJ01 | Rejection of invention patent application after publication |