CN113760358B - Antagonistic sample generation method for source code classification model - Google Patents
Antagonistic sample generation method for source code classification model Download PDFInfo
- Publication number
- CN113760358B CN113760358B CN202111003714.4A CN202111003714A CN113760358B CN 113760358 B CN113760358 B CN 113760358B CN 202111003714 A CN202111003714 A CN 202111003714A CN 113760358 B CN113760358 B CN 113760358B
- Authority
- CN
- China
- Prior art keywords
- conversion
- source code
- sample
- classification model
- current
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/70—Software maintenance or management
- G06F8/72—Code refactoring
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/40—Transformation of program code
-
- 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)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a countermeasure sample generation method oriented to a source code classification model, which comprises data preprocessing, extraction of a code conversion mode, selection of a candidate conversion mode, execution conversion, attack test and rewarding mechanism. Aiming at the defects and shortcomings of the existing countermeasure sample generation method, the invention extracts executable countermeasure operation aiming at the structural information of source codes, introduces the ideas of a Markov decision process and a time sequence difference algorithm, and adds the influence coefficient of the operation to guide the selection execution of the operation while executing the countermeasure operation, and continuously perfects the countermeasure sample generation method through decision learning, thereby realizing a quicker and more effective countermeasure sample generation method.
Description
Technical Field
The invention relates to a source code countermeasure sample generation technology, in particular to a countermeasure sample generation method oriented to a source code classification model.
Background
In order to meet the urgent needs of the current industry development, deep learning is increasingly applied to the field of source code classification, and corresponding processing and analysis, such as automatic marking of source codes in a program library according to functions, are beneficial to reuse of software programs in development and use. The deep learning model is used for completing automatic processing, analysis and generation of the program source code, so that the cost of system development, testing, operation, maintenance and the like of software can be greatly reduced. However, deep learning models present a very serious security risk, i.e. lack of robustness. And a method for solving this risk can be achieved by studying countermeasure learning. Challenge learning includes challenge attack and challenge sample generation.
Initially, the challenge sample generation was applied mostly in the image field and the natural language processing field. The image field is mainly focused on the structure of an attack graph, the nodes of the graph are considered to be fixed, and effective attack is realized by adding or deleting edges, so that a countermeasure sample is generated. The field of natural language processing is quite difficult to generate against samples than the field of images due to the discretization of language space, and word-level substitution or character-level inversion is often used to generate against samples. However, since source code has not only discrete features but also very strong structure, the challenge sample is subject to strict lexical, grammatical and grammatical rules.
The current method for generating the challenge sample in the field of source code classification mainly comprises a challenge sample generation challenge operation and a challenge sample generation operation execution strategy.
The antagonism sample generation antagonism operation is to perform antagonism operation on the source code to generate antagonism samples, and the antagonism samples are samples which are consistent with the original code semanteme but cause the target classification model to generate error result output after performing a series of operations on the source code. The operations of identifier replacement, synonymous expression replacement, insertion of dead code and the like are mainly performed on source code, and a reactance sample is generated by performing the countermeasure operations of the above types on the source code. However, according to research, the most widely-applied and latest function classification model is currently used, and the classification of source code functions is more based on capturing more grammar structure information, unlike the traditional text-based or label-based classification method. Research results also prove that more semantic information is usually contained in the program code structure, and the current generation mode still lacks processing and interference on the program code structure information.
The challenge sample generation operation execution policy is that when a challenge operation sequence required for generating a challenge sample is acquired, the operations need to be sequentially executed according to a set policy. Current challenge sample generation strategies are mainly accomplished by choosing a random method and an enumeration method. Random methods are easier to implement, but have a strong randomness and often fail to achieve more effective combat in a short period of time. The enumeration method has more comprehensive coverage when performing operations, but the calculation amount of the enumeration method also increases. Therefore, the enumeration without destination and direction can cause larger consumption of time and resources.
Therefore, the existing source code classification field fight sample generation method has the main defects that firstly, aiming at the working principle of the existing latest and most widely applied source code classification model, the highly discrete and structured characteristics of codes, the fight operation executed by the fight sample generation in the existing method is too simple and single; secondly, in the process of realizing the generation of the countermeasure sample, the execution strategy of the countermeasure sample generation operation in the existing method has low efficiency and large time and resource consumption.
Disclosure of Invention
The invention aims to provide an countermeasure sample generation method oriented to a source code classification model, which aims to solve the problems of excessively single countermeasure operation, low operation execution policy efficiency and high consumption in the existing generation method.
The invention is realized in the following way: a countermeasure sample generation method facing to a source code classification model comprises the following steps:
a. data preprocessing: for the original codes and the selected target classification model, firstly preprocessing the code fragments, namely processing the source codes into token sequences, or processing the source codes into an xml format by using an srcll tool;
b. extraction of code conversion modes: extracting a pair of source code samples < x, y > from the test set, and recording the initial state as S0; extracting conversion characteristics of each source code sample to obtain a corresponding executable conversion sequence A, limiting conversion operation in the whole iterative attack process of the current source code, and establishing a conversion operation value table Q (A) belonging to the current executable conversion sequence A, wherein the conversion operation value table Q (A) is used for recording the corresponding value of each conversion operation ai (ai E A,1 is less than or equal to i is less than or equal to k) so as to guide the selective execution of each operation;
c. selection of candidate conversion modes: selecting an execution conversion of the source code sample by using a ϵ -greedy algorithm with reference to a conversion operation value table Q (A), wherein the conversion operation value table Q (A) stores all executable conversions in a current executable conversion sequence A and continuously updated corresponding weights, and is used for measuring which conversion operation can obtain the highest expected rewarding value when the current source code is converted into a certain state;
d. performing conversion: c, executing the conversion operation selected in the step c on the source code sample x to generate a new converted code segment x';
e. attack test: d, sending the new code segment x' generated in the step d into a target classification model for attack test, if the classification result is different from the original function label, indicating that the attack is successful, and stopping the current iteration; otherwise, if the attack fails, repeating the step c and the step d;
f. rewarding mechanism: each attack test obtains and accumulates a reward value Rt, and updates a conversion operation value table Q (A) to guide the subsequent attack.
Further, after the data preprocessing in step a, the data is encapsulated by using pandas.
Further, the specific manner of selecting the switching operation in the step c is: setting a threshold ϵ, generating a random number before each operation selection, and if the random number is higher than the threshold ϵ, selecting to execute the conversion operation with the highest rewarding value in the current state; if the random number is below the threshold ϵ, any operation in the currently executable conversion sequence A is randomly selected for execution.
Further, the specific mode of step f is as follows: after the new code segment x' is sent into the original classification model to carry out attack test, obtaining a reward value Rt in the current state, if the sample obtained after the new conversion does not obtain a forward reward, discarding the current sample, and converting the sample to be attacked in the next iteration into a sample before the current conversion; all prize values are accumulated to obtain a cumulative prize value for the current sample sequence and the conversion operation value table Q (a) is updated with the prize value.
Further, the reward value is calculated by calculating a model classification probability label value.
Further, the forward rewards are models for helping misleading the target classification, and the forward rewards are obtained by calculating the reduction degree of the target label classification probability.
Aiming at the defects and shortcomings of the existing countermeasure sample generation method, the invention extracts executable countermeasure operation aiming at the structural information of source codes, introduces the ideas of a Markov decision process and a time sequence difference algorithm, and adds the influence coefficient of the operation to guide the selection execution of the operation while executing the countermeasure operation, and continuously perfects the countermeasure sample generation method through decision learning, thereby realizing a quicker and more effective countermeasure sample generation method.
The invention has the whole idea that aiming at the discretization and structuring characteristics of the source codes and combining the working characteristics of the existing latest source code classification deep learning model, the invention generates an reactance sample by extracting the abundant structural information of the source codes and executing structural conversion on the source codes. The idea of deep learning is introduced, and more effective conversion mode combinations are continuously learned, so that quicker and more effective countermeasure is realized in turn.
The method for generating the countermeasure sample for the source code classification model has the advantages that:
(1) Diversification of countermeasure operations: aiming at the structural characteristics of the source code and the current latest and most widely applied model classification working characteristics, various conversions aiming at the code structure are taken as countermeasure operations, and constraint of code semantics and grammar is still satisfied after the conversion is agreed so as to effectively generate countermeasure samples, and under the condition that the correct output of the source code is not changed, the DL model can be attacked to cause error results;
(2) Finite execution of operations: the method has the advantages that the resistance attack to the source code is realized by using a Markov decision process based on Q-learning, the problem of code search space explosion can be effectively solved, the search times are controlled within a limited time range, and the faster and more effective attack is realized;
(3) Validity of the selection strategy: the control problem of the whole experimental decision process is solved by using a time sequence difference algorithm, a rewarding mechanism is added, the conversion strategy is updated continuously by using the conversion value, and the attack is faster and more effective through continuous learning.
(4) The expandability is strong: the invention adopts a black box attack mode with the probe in the attack countermeasure scheme, does not need to use information such as a model architecture, internal parameters and the like for white box attack, only needs to obtain the value of model class label probability to guide the generation of an attack countermeasure example, can avoid the defect of excessive search times of the black box attack without the probe, can be conveniently and effectively applied to a plurality of deep learning models, and has good suitability and strong expandability.
Drawings
Fig. 1 is a flowchart of challenge sample generation.
Fig. 2 is a select transcoding stage development diagram.
Fig. 3 is an example display diagram of the present invention.
Detailed Description
The invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are intended to illustrate the invention and are not intended to limit the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
As shown in fig. 1, the method for generating an countermeasure sample for a source code classification model of the present invention includes the steps of:
step 1, data preprocessing: for the original code, the code fragment is preprocessed before all operations after execution. For the selected target classification model, the source code may be processed in two ways: one is serializing source code, processing the source code into a token sequence; another is to use a srcml tool to process the source code into xml format to facilitate countermeasure and restore. After data preprocessing, the data is encapsulated using pandas.
As shown in fig. 2, steps 2 and 3 of the present invention constitute the transcoding operations and selection phases.
Step 2, extracting a code conversion mode: as shown in part (1) of fig. 2, a pair of source code samples < x, y > is extracted from the test set, and its initial state is denoted as S0. And extracting conversion characteristics of each source code sample to obtain a corresponding executable conversion sequence A, limiting conversion operation in the whole iterative attack process of the current source code, and establishing a conversion operation value table Q (A) belonging to the current executable conversion sequence A, wherein the conversion operation value table Q (A) is used for recording the corresponding value of each conversion operation ai (ai E A, 1.ltoreq.i.ltoreq.k) so as to guide the selective execution of each conversion operation.
Step 3, selecting a candidate conversion mode: the conversion operation value table Q (a) is used to select the execution conversion of the source code sample by using a ϵ -greedy algorithm, and all executable conversions in the current executable conversion sequence a and the continuously updated corresponding weights are stored in the conversion operation value table Q (a) to measure which conversion operation can obtain the highest expected rewards value when the current source code is converted to a certain state.
The specific way of selecting the conversion operation is: setting a threshold ϵ, generating a random number before each operation selection, and if the random number is higher than the threshold ϵ, selecting to execute the conversion operation with the highest rewarding value in the current state; if the random number is below the threshold ϵ, any operation in the currently executable conversion sequence A is randomly selected for execution.
Step 4, performing conversion: as shown in fig. 3, for the source code sample x, the candidate conversion mode application selected in step 3 is executed to generate a new converted code segment x'.
Step 5, attack test: sending the new code segment x' generated in the step 4 into a target classification model for attack test, if the classification result is different from the original function label, indicating that the attack is successful, and stopping the current iteration; otherwise, judging that the attack fails, and repeating the step 3 and the step 4.
Step 6, rewarding mechanism: each attack test obtains and accumulates a reward value Rt, and updates a conversion operation value table Q (A) to guide the subsequent attack. The concrete mode is as follows: after the new code segment x' is sent into the original target classification model to carry out attack test, a reward value Rt in the current state is obtained, the reward value Rt is obtained by calculating the model classification probability label value, and the average reward value corresponding to each conversion operation in the new state can be obtained by using an incremental calculation mode, and the average reward value is related to the conversion operation executed and the times of conversion operation execution, so that the aim of the invention can be effectively realized. If the sample obtained after the new conversion does not obtain a forward rewarding, the current sample is abandoned, and the sample attacked by the next iteration is converted into the sample before the current conversion. All prize values are accumulated to obtain a cumulative prize value for the current sample sequence and the conversion operation value table Q (a) is updated with the prize value.
The forward rewards are used for helping misleading the target classification model, and the forward rewards are obtained by calculating the reduction degree of the target label classification probability.
And ending the current iteration when the attack is successful once or the set iteration times are reached. One sample repeatedly and iteratively acquires a plurality of value sequences in the continuously updated and learned operation value.
The invention mainly relies on executing structural conversion on source code, and can apply three types of conversion, namely control conversion, declaration conversion and API conversion. Taking the example of fig. 3 as an example, the countermeasure is implemented by extracting the structural information of the source code sample and selectively performing structural conversion on the code by adopting a selection strategy in the method.
Claims (6)
1. A method for generating an countermeasure sample for a source code classification model is characterized by comprising the following steps:
a. data preprocessing: for the original codes and the selected target classification model, firstly preprocessing the code fragments, namely processing the source codes into token sequences, or processing the source codes into an xml format by using an srcll tool;
b. extraction of code conversion modes: extracting a pair of source code samples < x, y > from the test set, and recording the initial state as S0; extracting conversion characteristics of each source code sample to obtain a corresponding executable conversion sequence A, limiting conversion operation in the whole iterative attack process of the current source code, and establishing a conversion operation value table Q (A) belonging to the current executable conversion sequence A, wherein the conversion operation value table Q (A) is used for recording the corresponding value of each conversion operation ai (ai E A,1 is less than or equal to i is less than or equal to k) so as to guide the selective execution of each operation;
c. selection of candidate conversion modes: selecting an execution conversion of the source code sample by using a ϵ -greedy algorithm with reference to a conversion operation value table Q (A), wherein the conversion operation value table Q (A) stores all executable conversions in a current executable conversion sequence A and continuously updated corresponding weights, and is used for measuring which conversion operation can obtain the highest expected rewarding value when the current source code is converted into a certain state;
d. performing conversion: c, executing the conversion operation selected in the step c on the source code sample x to generate a new converted code segment x';
e. attack test: d, sending the new code segment x' generated in the step d into a target classification model for attack test, if the classification result is different from the original function label, indicating that the attack is successful, and stopping the current iteration; otherwise, if the attack fails, repeating the step c and the step d;
f. rewarding mechanism: each attack test obtains and accumulates a reward value Rt, and updates a conversion operation value table Q (A) to guide the subsequent attack.
2. The method of generating an antagonistic sample for a source code classification model according to claim 1, wherein the data is encapsulated with pandas after preprocessing the data in step a.
3. The method for generating an countermeasure sample for a source code classification model according to claim 1, wherein the specific manner of selecting the conversion operation in the step c is: setting a threshold ϵ, generating a random number before each operation selection, and if the random number is higher than the threshold ϵ, selecting to execute the conversion operation with the highest rewarding value in the current state; if the random number is below the threshold ϵ, any operation in the currently executable conversion sequence A is randomly selected for execution.
4. The method for generating an countermeasure sample for a source code classification model according to claim 1, wherein the specific manner of step f is: after the new code segment x' is sent into the original classification model to carry out attack test, obtaining a reward value Rt in the current state, if the sample obtained after the new conversion does not obtain a forward reward, discarding the current sample, and converting the sample to be attacked in the next iteration into a sample before the current conversion; all prize values are accumulated to obtain a cumulative prize value for the current sample sequence and the conversion operation value table Q (a) is updated with the prize value.
5. The method of generating a challenge sample for a source code classification model of claim 4 wherein said reward value is calculated by calculating a model classification probability tag value.
6. The method of claim 4, wherein the positive rewards are directed to help misleading the target classification model by calculating the degree of degradation of the target label classification probability.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111003714.4A CN113760358B (en) | 2021-08-30 | 2021-08-30 | Antagonistic sample generation method for source code classification model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111003714.4A CN113760358B (en) | 2021-08-30 | 2021-08-30 | Antagonistic sample generation method for source code classification model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113760358A CN113760358A (en) | 2021-12-07 |
| CN113760358B true CN113760358B (en) | 2023-08-01 |
Family
ID=78791760
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111003714.4A Active CN113760358B (en) | 2021-08-30 | 2021-08-30 | Antagonistic sample generation method for source code classification model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113760358B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115080982B (en) * | 2022-06-24 | 2024-07-19 | 哈尔滨工业大学 | Combined anti-attack method for vulnerability detection model |
| CN117093997B (en) * | 2023-10-20 | 2024-02-20 | 广东省科技基础条件平台中心 | Code countermeasure sample generation method based on stable multi-arm slot machine |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325324A (en) * | 2020-02-20 | 2020-06-23 | 浙江科技学院 | Deep learning confrontation sample generation method based on second-order method |
| CN111737691A (en) * | 2020-07-24 | 2020-10-02 | 支付宝(杭州)信息技术有限公司 | Method and device for generating confrontation sample |
| WO2021109695A1 (en) * | 2019-12-06 | 2021-06-10 | 支付宝(杭州)信息技术有限公司 | Adversarial attack detection method and device |
| CN112949822A (en) * | 2021-02-02 | 2021-06-11 | 中国人民解放军陆军工程大学 | Low-perceptibility confrontation sample forming method based on double attention mechanism |
| CN112989346A (en) * | 2021-04-09 | 2021-06-18 | 鹏城实验室 | Countermeasure sample generation method and device, terminal device and storage medium |
| CN113221388A (en) * | 2021-06-17 | 2021-08-06 | 北京理工大学 | Method for generating confrontation sample of black box depth model constrained by visual perception disturbance |
-
2021
- 2021-08-30 CN CN202111003714.4A patent/CN113760358B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021109695A1 (en) * | 2019-12-06 | 2021-06-10 | 支付宝(杭州)信息技术有限公司 | Adversarial attack detection method and device |
| CN111325324A (en) * | 2020-02-20 | 2020-06-23 | 浙江科技学院 | Deep learning confrontation sample generation method based on second-order method |
| CN111737691A (en) * | 2020-07-24 | 2020-10-02 | 支付宝(杭州)信息技术有限公司 | Method and device for generating confrontation sample |
| CN112949822A (en) * | 2021-02-02 | 2021-06-11 | 中国人民解放军陆军工程大学 | Low-perceptibility confrontation sample forming method based on double attention mechanism |
| CN112989346A (en) * | 2021-04-09 | 2021-06-18 | 鹏城实验室 | Countermeasure sample generation method and device, terminal device and storage medium |
| CN113221388A (en) * | 2021-06-17 | 2021-08-06 | 北京理工大学 | Method for generating confrontation sample of black box depth model constrained by visual perception disturbance |
Non-Patent Citations (1)
| Title |
|---|
| 针对恶意代码分类模型的对抗技术研究;谷尚育;《中国科学技术大学 2020年硕士论文信息科技辑》(第01期);21-40 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113760358A (en) | 2021-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113760358B (en) | Antagonistic sample generation method for source code classification model | |
| CN112732919B (en) | Intelligent classification label method and system for network security threat information | |
| CN110008710B (en) | Vulnerability detection method based on deep reinforcement learning and program path instrumentation | |
| CN109542783B (en) | Extended finite-state machine test data generation method based on variable segmentation | |
| CN112732577A (en) | Evolution generation method for multi-task software test case | |
| CN113535399A (en) | NFV resource scheduling method, device and system | |
| CN116340952A (en) | Intelligent contract vulnerability detection method based on operation code program dependency graph | |
| CN112287645B (en) | Generation-type countermeasure network-based malicious PDF document generation method | |
| Petasis et al. | eg-GRIDS: Context-free grammatical inference from positive examples using genetic search | |
| CN116663019B (en) | Source code vulnerability detection method, device and system | |
| CN117714351A (en) | Network protocol fuzzy test method, system, equipment and medium | |
| CN118210706A (en) | Test case filtering method and device based on deep learning | |
| CN115392048B (en) | Constraint solving engine-based random number generation method with constraint | |
| CN110554952B (en) | Search-based hierarchical regression test data generation method | |
| CN116301875A (en) | Code semantic redundancy metric verification method based on triggerability of learning model | |
| CN112764791B (en) | Incremental update malicious software detection method and system | |
| KR102366153B1 (en) | Fast ensembel inference in machine learning for real-time application and machine learning method accelerated ensembel inference | |
| CN114840856A (en) | State-aware Internet of things trusted execution environment fuzzy test method and system | |
| CN114362992A (en) | Hidden Markov attack chain prediction method and device based on SNORT log | |
| CN108073679B (en) | Random pattern string set generation method and device in string matching scene and readable storage medium | |
| Psomopoulos et al. | A finite state automata based technique for protein classification rules induction | |
| CN112015641B (en) | Code odor detection method based on multi-granularity cascade forest model | |
| CN117435511B (en) | Flow monitoring software testing method and device based on graphic neural network and storage medium | |
| CN115114627B (en) | Malicious software detection method and device | |
| CN115481714A (en) | Deep learning vulnerability detection model confrontation sample generation method and system based on deep reinforcement learning |
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 |