阅读说明：本技术 一种基于深度学习的安卓应用管理系统及方法 (A kind of Android application management system and method based on deep learning ) 是由 王志强 李格菲 池亚平 张健毅 吕欣 邹潇湘 钱榕 张克君 卓子寒 于 2019-08-27 设计创作，主要内容包括：本发明涉及一种基于深度学习的安卓应用管理系统及方法,包括特征提取模块、深度学习模块以及分类管理模块。特征提取模块以安卓应用为输入,使用静态提取方式提取特征形成特征文档。深度学习模块使用卷积神经网络作为学习模型。一方面对特征文档与标识值组成的样本集进行训练与学习,得到学习成熟的模型；另一方面以特征文档作为输入,使用学习成熟的模型,准确的分类管理安卓应用,提高应用商店的管理效率。(The present invention relates to a kind of Android application management system and method based on deep learning, including characteristic extracting module, deep learning module and Classification Management module.Characteristic extracting module is applied with Android as input, is extracted feature using static extracting mode and is formed feature documents.Deep learning module uses convolutional neural networks as learning model.On the one hand the sample set that feature documents are formed with ident value is trained and is learnt, obtain the mature model of study；On the other hand using feature documents as input, using the model that study is mature, accurate Classification Management Android application improves the efficiency of management of application shop.)

1. a kind of Android application management system based on deep learning characterized by comprising characteristic extracting module, depth Practise module and Classification Management module, in which:

Characteristic extracting module, the characteristic extracting module, which is configured as applying with the multiclass Android with different classes of ident value, is Input extracts multiclass feature using static extracting mode and forms feature documents；

Deep learning module, the deep learning module are configured with convolutional neural networks to by the feature documents and class The sample set of other ident value composition is trained, and obtains the mature model of study；Simultaneously using the feature documents as input, make Use the model of the study maturation as final Classification Management model；

Classification Management module, the Classification Management module use browser/server framework, and wherein server end is configured as leading to Cross the classification that the mature model of the study for calling the characteristic extracting module and deep learning module to obtain is applied as Android Administrative model；Browser end is configured as carrying out uploading a large amount of new Android applications towards Android application shop, passes through server The model at end carries out Classification Management to Android application.

2. a kind of Android application management method based on deep learning, it is characterised in that: the following steps are included:

The first step, acquisition largely have the Android application of classification logotype, and batch extracting static nature forms feature documents；

Second step is trained the sample set being made of feature documents and classification logotype value using convolutional neural networks, obtains Learn mature model；

Third step is uploaded onto the server Android using batch by browser, and the mature model of server calls study carries out Classification Management.

3. the Android application management method according to claim 2 based on deep learning, it is characterised in that: in the first step, The extraction static nature and form feature documents the following steps are included:

3.1. Android application is decompressed, file AndroidManifest.xml is obtained；

3.2. AndroidManifest.xml file is parsed using AXMLPrinter2 and TinyXml tool, obtains Android application Permission feature, software and hardware feature and intent features as static nature；For each Android application, included by it Static nature extracts in text document as feature documents.

4. the Android application management method according to claim 2 based on deep learning, it is characterised in that: in second step, The convolutional neural networks are made of embeding layer, convolutional layer, pond layer and full articulamentum, are had the feature that

4.1. embeding layer: traversing all feature documents and obtain non-duplicate static nature, maps that the expression of low-dimensional vector, embedding Entering dimension is 300, is indicated thus to obtain the corresponding vector of each static nature；The feature documents and its mark that each Android is applied Knowledge value is input to the embeding layer of convolutional neural networks, and the static nature for being included by feature documents is converted to vector expression, obtains Two-dimensional eigenmatrix；

4.2. convolutional layer: being connected in parallel 4 convolutional layers conv1, conv2, conv3 and conv4 after embeding layer, convolution kernel Size is respectively 2x300,3x300,4x300 and 5x300, and the number of convolution kernel is 64；

4.3. pond layer: be separately connected after four convolutional layers maximum pond layer pooling1, pooling2, pooling3 and Pooling4, each pond layer obtain 64 optimal characteristics；

4.4. full articulamentum: being sent into softmax classifier by full articulamentum after the layer of pond, carries out Classification Management.

Technical field

The present invention relates to mobile terminal safety technical field, in particular to a kind of Android application management based on deep learning System and method.

Background technique

For Android as presently most popular intelligent movable operating system, equipment and number of users are huge, rich using type Richness, safety have received widespread attention.Classification that Android application shop is specified generally according to application developer or by point The description that analysis application developer provides classifies to application.But this process is easy to by the developer of malicious application Manipulation applies addition to be easier to the information class application interface by audit to escape detection, such as by no qualification finance class.Furthermore with The explosive growth of Android number of applications, how rapidly and accurately automatically to be classified to Android application, the raising efficiency of management It is vital.But it is less for the research of Android application class management at present.In maximally related work, Shabtai Et al. Classification Management is carried out to tool-class and the application of game class Android using static analysis and machine learning algorithm；Wang Wei etc. People carries out Classification Management to the application of 24 class Androids also with Static Analysis Technology, using a variety of machine learning algorithms.But machine Device learning algorithm is expressed complicated function and classification problem and is limited, and generalization ability is restricted, and accuracy rate is lower.

Summary of the invention

Technology of the invention solves the problems, such as: overcoming the deficiencies of the prior art and provide a kind of Android based on deep learning and answers With management system and method, Android application feature is extracted using static extracting mode, to ensure Classification Management efficiency；Android is answered Feature documents processing is two-dimensional eigenmatrix, and form turns to " picture ", using what is done well in picture recognition field Convolutional neural networks algorithm carries out accurate Classification Management to the application of a large amount of Androids.

Technical solution of the invention: Android application management system and method based on deep learning are wrapped as shown in Figure 1 It includes:

Characteristic extracting module, the characteristic extracting module are configured as answering with the multiclass Android with different classes of ident value With to input, uses static extracting mode to extract multiclass feature and form feature documents；

Deep learning module, the deep learning module are configured with convolutional neural networks to by the feature documents It is trained with the sample set of classification logotype value composition, obtains the mature model of study；Simultaneously using the feature documents as defeated Enter, uses the model of the study maturation as final Classification Management model；

Classification Management module, the Classification Management module use browser/server framework, and wherein server end is configured For what is applied by the mature model of the study for calling the characteristic extracting module and deep learning module to obtain as Android Classification Management model；Browser end is configured as carrying out uploading a large amount of new Android applications towards Android application shop, passes through clothes The model at business device end carries out Classification Management to Android application.

The present invention is based on the Android application management systems of deep learning to realize that steps are as follows:

1. acquisition largely has the Android application of classification logotype, batch extracting static nature forms feature documents；

1.1. Android application is decompressed, file AndroidManifest.xml is obtained；

1.2. AndroidManifest.xml file is parsed using AXMLPrinter2 and TinyXml tool, obtains Android Permission feature, software and hardware feature and the intent features of application are as static nature；For each Android application, it is wrapped The static nature contained extracts in text document as feature documents；

2. being trained using convolutional neural networks to the sample set being made of feature documents and classification logotype value, learned Mature model is practised, the convolutional neural networks are made of embeding layer, convolutional layer, pond layer and full articulamentum；

2.1. embeding layer: all feature documents are traversed and obtain non-duplicate static nature, map that low-dimensional vector table Show, insertion dimension is 300, is indicated thus to obtain the corresponding vector of each static nature.The feature documents that each Android is applied And its ident value is input to the embeding layer of convolutional neural networks, the static nature for being included by feature documents is converted to vector table Show, obtains two-dimensional eigenmatrix.

2.2. 4 convolutional layers conv1, conv2, conv3 and conv4, convolution convolutional layer: are connected in parallel after embeding layer The size of core is respectively 2x300,3x300,4x300 and 5x300, and the number of convolution kernel is 64.

2.3. pond layer: be separately connected after four convolutional layers maximum pond layer pooling1, pooling2, pooling3 and Pooling4, each pond layer obtain 64 optimal characteristics.

2.4. full articulamentum: being sent into softmax classifier by full articulamentum after the layer of pond, carries out Classification Management.

3. Android is uploaded onto the server using batch by browser, the mature model of server calls study is divided Class management.

The advantages of the present invention over the prior art are that:

(1) present invention is one kind effectively supplement to Android application management method.It is answered using static extracting mode from Android It is permission feature, software and hardware feature and intent features respectively, compared to dynamic analysis, consumption is more with three category features of middle extraction Few resource and time.Because application scenarios of the invention are mainly Android application shop, merely with Static Analysis Method Feature is extracted, to ensure Classification Management efficiency.In addition, the system has certain expansion, it can be according to Android application shop Free adjustment is carried out for the division of management category, while more features can also be added, more accurately to carry out Android Good basis is laid in application class management.

(2) two-dimensional eigenmatrix is converted by the feature that each Android is applied, is visualized as " picture ", is incorporated in image Identification aspect show outstanding convolutional neural networks model can more efficient learning characteristic, it is quasi- to reach higher Classification Management True rate and faster Classification Management speed.

Detailed description of the invention

Fig. 1 is the flow chart of present system；

Fig. 2 is the structure of the convolutional neural networks in present system.

Specific embodiment

The present invention is described in detail with reference to the accompanying drawing

As shown in Figure 1, the present invention is based on the Android application management systems of deep learning by characteristic extracting module, deep learning Module and Classification Management module composition.

Illustrate so that millet application shop application class manages as an example below, entire to realize that process is as follows:

(1) application of multiclass Android is downloaded from millet application shop crawler, including game class is applied, books reading class is applied, Audio-visual audiovisual class application, the application of chat social category, the application of sports class, the application of Domestic News class, the application of fashionable shopping class, gold Melt the application of financing class and photography and vedio recording class application, and manual correction is carried out to the application of wherein classification error, by above nine class Application order number is 1~9, as classification logotype value；

(2) decompress Android application, obtain file AndroidManifest.xml, using AXMLPrinter2 and TinyXml tool parse AndroidManifest.xml file, obtain Android application permission feature, software and hardware feature and Intent features are as static nature, i.e. label<uses-permission>,<uses-feature>,<intent-filter>in Content.For each Android application, the static nature for being included by it is extracted in text document as feature documents；

(3) sample set being made of feature documents and classification logotype value is trained using convolutional neural networks, is obtained Learn mature model.Convolutional neural networks include embeding layer, convolutional layer, pond layer and full articulamentum.

(3.1) embeding layer: all feature documents are traversed and obtain non-duplicate static nature, map that low-dimensional vector table Show, insertion dimension is 300, is indicated thus to obtain the corresponding vector of each static nature.The feature documents that each Android is applied And its ident value is input to the embeding layer of convolutional neural networks, the static nature for being included by feature documents is converted to vector table Show, obtains two-dimensional eigenmatrix.

(3.2) 4 convolutional layers conv1, conv2, conv3 and conv4, volume convolutional layer: are connected in parallel after embeding layer The size of product core is respectively 2x300,3x300,4x300 and 5x300, and the number of convolution kernel is 64.

(3.3) maximum pond layer pooling1, pooling2, pooling3 pond layer: are separately connected after four convolutional layers And pooling4, each pond layer obtain 64 optimal characteristics.

(3.4) full articulamentum: being sent into softmax classifier by full articulamentum after the layer of pond, carries out Classification Management.

(3.5) 80% is randomly selected in the application of all Androids as training sample, wherein 10% is used as verifying sample, 20% is remaininged as test sample.In each training process, the accuracy rate and penalty values of observation verifying sample, when penalty values no longer drop Deconditioning when low obtains the mature model of study, as final Classification Management model.

(3.6) laboratory test results of the invention are as shown in table 1, the Classification Management model obtained using deep learning module 20% test sample is tested, six seed type Android application all classifications are correct as the result is shown, three types Android application In respectively have an application class mistake.Demonstrate the classification validity of this system.Subsequent a large amount of Android applications can be carried out quick Accurate Classification Management.

Table 1

(4) Android application shop can be uploaded onto the server Android using batch by browser, server calls It practises mature model and carries out Classification Management.