Android Malware Detection Using Fine-Grained Features

,


Introduction
Smartphones have become an integral part of our day-to-day life.New data for December 2018 shows that Android remains the most popular mobile operating system, with a worldwide market share of 75.16% [1].With over one million Android applications in major app stores, applications such as WeChat, TikTok, and mobile banking applications are used in our daily life and continue to play an increasingly important role.Most of these applications have access to users' private information such as their location, credit card, and contact information.Almost all applications access the users' private data, although this provides users with better personalized services [2].It may also result in information leakage of private data and economic loss [3].Further, Android malicious applications keep emerging endlessly, and this security issue has gained increasing attention in the industry and academic fields [4].
A large body of research against Android malware has been proposed.Currently, static analysis and dynamic analysis are the two main types of detection methods.Each approach has its merits and shortcomings.e static analysis methods such as Kirin [5], PApriori [6], and DREBIN [7] analyze applications without executing the program requiring low overheads.However, the methods cannot defend against antidecompiling and obfuscation.On the contrary, dynamic analysis methods such as TaintDroid [8] and VetDroid [9] execute the application in real time to detect malware, but it is difficult to acquire all the execution paths.With malware being rapidly evolving, the machine learning method is used to perform Android malware detection.Consequently, gathering features that better represent malicious behavior as the features of machine learning is beneficial to improve the performance of malware detection.
By itself, Android has several security mechanisms in its different layers.e permission mechanism applied in the application layer is an important defence mechanism to protect sensitive resources on the Android platform.Applications must declare dangerous permissions to access sensitive data [10,11].Several studies have investigated Android malicious applications based on the declared permissions, using permission-based methods [5][6][7][12][13][14].Although these methods avoid high overhead, they consider the declared permissions as the features of machine learning, which cannot truly reflect the difference between benign applications and malicious applications.us, they cannot detect malicious applications that declare only a few, or dangerous permissions, which are also always declared by benign applications.
Compared with permissions as features, application programming interfaces (APIs) represent the entire picture of application behavior provided by the Android system [15].DroidAPIMiner [16] exploits data flow analysis to extract the numbers of APIs used in malicious applications and benign applications to analyze the difference between them, which is similar to these methods in [17,18].In general, the API feature set contains a very large number of features, and therefore, detection methods need extra time to extract the API features and to train detection models.It is worth noting that there is a corresponding relation between permissions and APIs.
In this paper, we present FDP, a lightweight Android malware detection method to mine hidden patterns of malware.According to previous studies, there is considerable difference between malicious applications and benign applications in terms of the declared permissions [19][20][21][22].In addition, some dangerous permissions declared for different components reflect the purpose of the developers [23,24].It can be used to distinguish different purposes of the same dangerous permission.erefore, FDP exploits static analysis to collect more fine-grained and representative features from AndroidManifest.xml and decompiled code; these features include fine-grained permissions, intent filter, and the code features of malicious applications.ese features can represent the difference between benign applications and malicious applications.
Experiments with 1700 benign applications from Xiaomi markets and 1600 malicious applications demonstrate the effectiveness of FDP, which achieves a TP rate of 94.5% and only requires 15.052 s to analyze an application on average.e main contributions of our work can be summarized as follows: (1) We propose a new method to perform Android malware detection based on fine-grained permission mechanism, which represents the difference between malicious applications and benign applications as the features of machine learning, including the information of the dangerous permissions used in the components for the first time.
(2) We present a thorough study on the permissions frequently used by malicious applications and evaluate the importance of permission features used to classify malicious applications and benign applications.(3) In terms of efficiency, FDP uses static methods to gather all features and analyzes an application in a reasonable time.(4) e experiments demonstrate that FDP is more effective for detecting more malware families and malicious applications that declare few dangerous permissions or dangerous permissions as those declared by benign applications.
e remainder of this paper is organized as follows.Section 2 introduces the features of malicious applications.Section 3 covers the detection framework, including feature selection, feature extraction, and the fine-grained handle on the features.Section 4 introduces the key points for backtracking.Section 5 discusses the experiments and the results.Section 6 describes related work, and we conclude the paper in Section 7.

The Features of Malicious Applications
According to Google's definition, permissions are divided into several protection levels: normal, signature, dangerous, and special permissions.Normal permissions have very little risk to the user's privacy.Signature permissions are granted by Android during installation.Dangerous permissions refer to resources that involve the user's private information, which are shown in Table 1.Special permissions do not work like normal and dangerous permissions, which are particularly sensitive; therefore, most applications do not use them [25].Malicious applications must exploit dangerous permissions to execute malicious behavior.According to the previous work, the research studies about dangerous permissions play an important role for the detection of malicious applications.
Moreover, malicious applications have some code features.AndroMalShare [26] holds 86,798 malicious applications and presents statistical information of the samples, which includes dynamic loading, native code, and reflection, as shown in Figure 1.
e reasons why malicious applications have the code features are provided as follows: Dynamic loading: Android supports applications to load additional binaries at run time by dynamic loading.In order to evade static detection, malicious application developers separate the core functionality of the application into independent libraries and load them dynamically [27].erefore, malicious applications always employ dynamic loading to hide malicious behavior.Native code: many developers use the native code as part of their applications to improve the execution efficiency of the code and to increase the difficulty of decompilation.Afonso et al. [28] estimate that 37% of all Android applications have the native code.Most malware is restricted to the detection of bytecode; therefore, malware developers use the native code to implement malicious behavior lest the code of the application is reversed, which means that static analysis cannot analyze the applications including the native code completely [29].Reflection: since Android applications are composed of the Java code and native code, they have the ability to leverage the reflection mechanism.Malicious applications leverage reflection to invoke APIs corresponding 2 Scientific Programming to dangerous permissions to perform malicious behavior, in an attempt to evade static detection [30].

Detection Methodology
3.1.Overall Framework.In order to extract features that can better represent application behavior, FDP exploits static analysis to extract key features by decompiling applications and analyzing them using the optimal algorithm of machine learning.e process is illustrated in Figure 2 and the details are provided as follows: (a) Static analysis: FDP employs the apktool to decompile the application and acquires different feature sets from the AndroidManifest file and the decompiled source code.(b) Embedding in vector space: all extracted features are mapped to a vector space that can be analyzed as features in machine learning.If the application has the features, the corresponding dimensions are set to 1. Instead, the corresponding dimensions of the other features are set to 0. In order to prevent the problem of overfitting, we take the action of "RemoveDuplicates" to remove the same instances.(c) Optimal classifier: we use J48 (decision tree algorithm), K-Nearest Neighbor (KNN), Naïve Beyesian (NB), and Support Vector Machines (SVM) to analyze the same training set and choose the classifier with the best indicators as the optimal classifier for FDP.

Fine-Grained Permissions.
e permissions mechanism restricts access to a series of critical APIs [31]. is paper refers to the APIs corresponding to dangerous permissions as sensitive APIs.Several studies choose APIs as the features of machine learning.e classification information is relatively rich, provided that all sensitive APIs are selected as features.However, it is time consuming to process with huge resource consumption and there must exist redundant and meaningless features.Nevertheless, the premise of invoking sensitive APIs is that the applications must declare permissions in the AndroidManifest file.For example, applications must declare the SEND_SMS permission to invoke sendTextMessage() and sendMultipartTextMessage() to send messages.erefore, FDP selects permissions instead of APIs as the features of machine learning to ensure it is lightweight.
In terms of how to select permissions, it is not suitable to select permissions that are commonly used by both malicious applications and benign applications.If all permissions are used, it will inevitably lead to a "curse of dimensionality."As shown in the previous work [19], malicious applications tend to declare dangerous permissions more often than benign applications.us, our research method considers the permissions that are frequently used by malware as the object of our study and extracts the features from the following aspects.
Declared permissions and intent: applications invoking sensitive APIs must declare the corresponding permissions in the AndroidManifest file.We extract these declared permissions after decompiling the application.Activities, Service, and Broadcast Receivers are activated by intent and they register their type of intent using intent-filters in the AndroidManifest file.Malicious applications can exploit the announcement from the Android system to trigger malicious behavior.erefore, our method extracts the intent that is always used in malicious applications as features, e.g., BATTERY_CHANGE_ACTION and SMS_RECEIVED.
Unused permissions and root privilege: some applications declare dangerous permissions; however, there are no mapped APIs in the Smali code of the application because dangerous permissions are used in dynamic loading or abused by the developers.
Our method first extracts dangerous permissions listed in the AndroidManifest file.en, we traverse the decompiled source codes based on the relation between the permissions and the APIs provided by PScout [32], to look for declared permissions for which the mapped APIs are not invoked.Such dangerous permissions are defined as unused  permissions.If there exists sensitive APIs without mapped dangerous permissions in the AndroidManifest file, we suspect that the application is probably attempting to gain root privileges.Because the root privileges give the application absolute control over the device, it can execute any malicious behavior without permissions [33].
Permissions used in components: permissions used in components are the bright spot of our method.In this study, the sensitive APIs are used as the starting point to backtrack the generation of the call graph.According to the relation between the permissions and the APIs provided by PScout, our method scans the Smali code and looks for the mapped APIs.en, with these APIs as a starting point, the process for the backtracking of parent function is iterated until the component can be identified through the files where a parent function is in.
e specific steps of the generation process of backtracking are as follows: (1) Traverse all the Smali files in the folder after decompiling, and find out the sensitive API, such as Landroid/ telephony/TelephonyManager;>getDeviceId()Ljava/lang/ String.en, look for the function that contains the sensitive API, which is the nearest function from the bottom to the top.Taking Figure 3 as an example, the function containing the sensitive API is getDid().
(2) Generate a new parent function according to the sensitive API and the class it is in.We take the new parent function as the starting point, traverse all the Smali files in the folder again and find out the function which calls the newly generated parent function.Taking Figure 3 as an example, it means that we should find out where getDid() is called, i.e., Lcom/qq/e/v2/managers/status/DeviceStatus;->get-Did()Ljava/lang/String.
(3) Iterate the methods of steps 1 and 2 until we get the component information where the parent function is located.e iteration result of getDid() is shown in Figure 4. e statement of ".super" displays the component information.e ".super Landroid/app/Activity" in the Smali file indicates that the parent function called Landroid/telephony/TelephonyManager;->getDeviceId() Ljava/Lang/String belongs to the Activity component.

Native Code and Reflection.
After decompilation, our method checks whether the application has the lib folder that stores the libraries.If the lib folder exists, the feature of the native code is set to 1.In order to alleviate the absence of reflection, FDP extracts APIs related to reflection, such as Class.forName().e feature enables us to know that the hidden code is executed.erefore, the feature of reflection is set to 1 or 0 depending on the code related to the reflection.

The Key of Backtracking
e feature of permissions used in components contributes a lot for FDP to improve the detection performance.erefore, 4 Scientific Programming our method must backtrack the parent function calling the sensitive API to get the component information.Moreover, the key to the whole backtracking process is that the building process of the call graph cannot be interrupted.However, owing to the communication between components, callback, and other reasons, Android applications have implicit calls, which cause the interruption of a call graph depending on the conventional method.Some situations and solutions are presented as follows.
4.1.Annotations.Annotations are a language feature of Java widely used in the development of Android applications [34].ere are two types of annotations: the dalvik.Annotation package is not open to the public, which is only used for the core library and code testing, and it is under the Dalvik\src\main\Java\dalvik\annotation directory; the other one is android.Annotation and the corresponding annotation declaration is under the framework \base\-core\Java\android\annotation directory.Annotation often appears in the process of backtracking.Taking the app "net.maxicom.android.snake"as an example, the code snippet is shown in Figure 5.When the detection backtracks the parent function, FDP encounters the annotations code, in which "EnclosingMethod" specifies the scope of its own class."Method" shows that the annotation acts on a function, and the value tells us that it is located in onCreate () of SnakeService.erefore, FDP opens the SnakeService file, where the second line is ".Super Landroid/app/Service." us, it can be inferred that the sensitive API is in the Service component.

Multithread Communications.
Android is messagedriven, and the essence of the message-handling mechanism is that one thread opens a loop to continuously listen to and process messages sent by other threads in turn.If a new thread is created to operate on the main thread, the system will throw an exception.erefore, an asynchronous callback mechanism Handler is provided in the Android system for communication between threads [35].
As shown in Figure 6, the UI thread is the main thread.e system initializes a Looper object and also creates a MessageQueue associated with it.e Message is the object that the Handler receives and processes.e MessageQueue is the Message queue.Each UI thread can only have one Looper, continually pulling the Message out of the Messa-geQueue and distributing it to the corresponding Handler.
When Looper gets a new message, handleMessage() is called to process the new message.is Android message-handling mechanism causes the call graph to break, which leads to the failure of backtracking.erefore, our method constructs an instance initialization method that has the special name "<init>," which is supplied by a compiler.en, we use the method as the new starting point of backtracking.
Taking "ServiceCommunication1.apk" as an example, we backtrack the parent function calling sensitive API until the parent function is "handleMessage," as shown in Figure 7(a).

Process and read.
A process may contain several threads, which can be used as the basic unit of independent operation and independent scheduling.Because threads are smaller than processes and do not own system resources, their scheduling costs are low, which effectively improves the concurrent execution of programs between multiple programs of the system.
ere are two ways to implement threads in an Android system.One is to extend the "java.lang.threads,"rewrite "run()," and use multiple threads to complete their tasks separately.e other is to implement the Runnable interface and instantiate the thread class.e difference is that the threads created with the Runnable interface can share resources when multiple threads have access to the same resource, whereas the threads created by inheriting the thread class have their own resources.
When the parent function of the sensitive API is "run()," it is necessary to check the keywords ".Super" and ".Implement" in the Smali file.If there exists ".super Ljava/lang/ read" or ".Implement Ljava/lang/Runnable," we need to traverse all the Smali files again, instantiate the thread and use the function that starts the thread as the starting point for backtracking, and search upward for the parent function.Note that the thread class has the methods of "run()" and "start()," whereas the implementation of the Runnable interface only has the "run()" method.
Taking "net.maxicom.android.snake"as an example, the application invokes location/LocationManager;->requestLocationUpdates(Ljava/lang/String; JFLandroid/ location/LocationListener)V corresponding to the ACCESS_FINE_LOCATION permission, and its parent  Lnet/max-icom/android/snake/Snake-Service$1;->run(V), where the second line of Snake-Service$1 shows ".super Ljava/lang/ read" in the "net/ maxicom/android/snake/SnakeService$1" file.Provided that the method constructs its parent class function by class, method, and parameter as previously addressed, the backtrack fails.We take Lnet/maxicom/android/snake/ SnakeService$1;-> start (V) as the new starting point for backtracking.Eventually, after scanning and matching, the parent function is Lnet/maxicm/android/snake/ SnakeService;->onCreate(), and the file is located in net/ maxicom/android/snake/snakeService where the second line shows that the sensitive API is called in the Service component, as shown in Figure 8.

Experiment Setup
Step 1: build a sample dataset.
e malicious application samples come from MalGenome and Andro-MalShare [26].MalGenome is widely used by many researchers which help us compare the results with related approaches.As a remedy, AndroMalShare provides the latest malware samples which can be classified by malware families.We collect 1600 malicious applications as samples, which are classified according to Android malware families.advantage of the dataset is that we can select specific Android malware families to build the detection model and examine the ability of the detection model detecting unknown applications from the other Android malware families.In order to ensure the balance of the data sets, we employ the crawler program to download 1700 benign applications from the Xiaomi App Store.To further check benign applications and malicious applications, we validate these applications with Virustotal, which is the website that analyzes suspicious files and URLs to detect types of malware, automatically sharing them with the security community.We select 20 malware families shown in Step 2: decompile the application using the apktool tool to get the Smali code and AndroidManifest file.en, we extract all features from the Smali code and AndroidManifest file.After further analysis, the finegrained features are acquired.
Step 3: use the program written in Python to unify the features extracted in Step 2, and unify it into the arff format, which is convenient for machine learning by WEKA tools.
Step 4: compare the classification results of two different sets of permission features to select the features that can better represent the malicious behavior and exploit different machine learning algorithms to learn the data and select the optimal classifier according to the TP Rate, F-measure, and Receiver Operating Characteristic curve (ROC) area.
All experiments are carried out on a machine with a memory of 16 GB and Intel (R) Core(TM) i7-4720hq 2.60 GHz processor.

Evaluation Metrics.
In this work, TP Rate, F-measure, and ROC area are employed to evaluate the performance of the detection model.As malicious applications are positive samples and benign applications are negative samples in our evaluation, we present four types of values.t pos is the number of malicious applications correctly identified as malicious applications; f neg is the number of malicious applications incorrectly identified as benign applications; t neg is the number of benign applications correctly identified as benign applications; and f pos is the number of benign applications incorrectly identified as malicious applications: ( In these metrics, F-measure is an indicator referring to precision and recall.
e ROC area is one of the most important evaluation metrics for checking any classification model's performance.e higher the ROC area, the better the classification.

Selection of Features.
In order to achieve a better detection performance, choosing informative, discriminating, and independent features is a crucial step for classification.We design two experiments to compare the permission sets and use the testing set to validate.Experiment Ι: select all dangerous permissions as features of machine learning; Experiment II: select the top 20 used permissions of 86798 malware samples collected by AndroMalShare as the features of machine learning, as shown in Figure 9.
Experiment Ι chooses 927 malicious applications and 1000 benign applications that do not contribute to the detection model as a testing set.e result of dangerous permissions as features of machine learning is shown in Table 3. KNN and SVM are relatively better than other classifiers, and they have the highest TP Rate in malware detection.However, the performances of the four classifiers are not very satisfactory.
In Experiment II, we use the same training set to build the detection model and analyze the same testing set by the same algorithms.e only difference is that the features are the top 20 used permissions of 86,798 malicious applications collected by AndroMalShare.e experimental result is shown in Table 4.
It can be seen from the results that the SVM algorithm is the best in the TP Rate, reaching 90.9%.To compare the results in Tables 3 and 4, similar performances are observed for the same testing set and classifier.It is difficult to evaluate which permission set is better; however, we can identify the features which are more important for classification depending on the information gain of the features.e two sets of results are shown in Tables 5 and 6, respectively.
According to the ranked features and the actual usage status of permissions, we choose the 24 permissions as features from the two sets, as shown in Table 7.And we find that the permissions, that is, the requirement of executing malicious behavior may not be the significant like INTERNET permission, which is used frequently by malware and benign applications.en, we employ the selected permission features as features of machine learning to learn the training set and analyze the same applications from the test set with the same classifiers.e result is shown in Table 8, with all indicators being superior to those in the above experiments.In these algorithms, the highest TP Rate achieved is 93.8% and the Fmeasure is the best with the J48 classifier.

FDP.
rough further analysis on classifier errors, we find that most of these applications are wrongly classified as they declare dangerous permissions that are frequently declared by malware applications and benign applications, especially, READ_PHONE_STATE, ACCESS_COARSE_-LOCATION, ACCESS_FINE_LOCATION, SEND_SMS, and READ_SMS.Although these dangerous permissions are exactly same, the usage of these permissions is different between benign applications and malicious application.For example, the information gain of READ_PHONE_STATE is relatively low, but malicious applications often employ it in service component in order to avoid any users' attention which is different from benign applications.
Based on the previous work [23,24], we further refine the above permissions and extract the information of  FDP also uses the same data sets and classifiers for analysis.e result is shown in Table 10.Every algorithm has better performance than before.In these classifiers, J48 is the best and its TP Rate and F-measure are 0.945.It demonstrates that the features of FDP can better represent the malicious behavior.
Besides, the results of FDP are also shown as a ROC area in Figure 10.e area below the ROC curve is the value of the area under the curve (AUC).e larger the area, the better the classification.e results are obvious that these four classifiers have good performance in detecting malicious applications and all ROC area are over 0.9.

5.5.
e Evaluation of Dangerous Permissions Used in Component.In order to evaluate the importance of the feature of dangerous permissions used in components, we collect malicious applications from Android malware families, such as FakePlayer, DroidCoupon, TapSnake, and Plankton.
ese applications only declare few dangerous permissions such as READ_PHONE_STATE, ACCESS_-COARSE_LOCATION, and READ_SMS which are frequently used by benign applications.is means there is not enough available information for classification if the detection methods are based on the declared permissions.
For comparison, we employ two sets of the permission features.e features in Table 7 are taken as one of two sets, and the other is that the features of dangerous permissions applied in components replace the declared permissions based on the features of Table 7, which involve READ_ PHONE_STATE, ACCESS_COARSE_LOCATION, ACCESS_ FINE_LOCATION, and READ_SMS.For example, READ_ PHONE_STATE is replaced by A-READ_PHONE_STATE, B-READ_PHONE_STATE, and S-READ_PHONE_STATE.
By testing 63 malicious applications and 70 benign applications that declare similar permissions, we perform 10-fold cross-validation using SVM algorithm.e accuracy of detection is 77.8% with the declared permissions, while the accuracy of detection reaches 92.5% with the finegrained permissions.e scale of two data sets is relatively small because there are too few eligible benign applications.
In addition, we use information gain to evaluate the importance of features including the fine-grained permissions.As shown in Table 11, there are three features of dangerous permissions used in components ranked in the top fourteen of permissions.
e result demonstrates the feature of permissions used in components contributes to the classification.

Detection of Unknown Malware Families.
In order to examine the prediction ability of FDP, we collect the malware families as the unknown malware families shown in Table 12.ese applications are not in the training set for constructing the detection model, which include the malware families that PApriori and DREBIN cannot detect.
We use three detection models to identify the same applications from unknown malicious families.e three detection models employ dangerous permissions as the features of machine learning (i.e., Experiment Ι), the top 20 used permissions of malicious applications declared as the features of machine learning (i.e., Experiment II), and FDP, respectively.e experimental results show that FDP can accurately detect more malicious families, especially with few dangerous permissions, which are shown in Figure 11.For instance, all the malicious     PApriori detects 87% of the malware samples.However, when malicious applications declare less dangerous permissions as benign applications, PApriori fails.PApriori lists malware families from MalGenome undetected by itself, such as Asroot, FakePlayer, DroidCoupon, DroidDeluxe, Tapsnake, and Plankton.And that is why PApriori has obvious limitations in detecting benign applications.FDP can detect all the above malicious families because of the fine-grained selected dangerous permissions, which represents the difference between benign applications and malicious applications.DREBIN detects malware with the accuracy of 94%, but it cannot detect malicious applications from the Gappusin family because there are too few malicious features to identify the sample.We employ applications from Gappusin family as unknown malicious applications and the TP Rate gets 100%.Meanwhile, there is the problem of unbalanced data in the data sets of DREBIN in which it contains 123,453 benign applications and 5,560 malicious applications.If DREBIN chooses the balanced data sets, the accuracy may improve.
DroidEnsemble employs string features such as permissions and intent and structural features to detect malware applications.e size of the dataset and the composing proportion of benign applications and malicious applications are similar to FDP, and the accuracy reaches 98.4%.It is worth noting that structural features may perform poorly in detecting unknown malicious families because it relies heavily on previous identified malicious families.erefore, it has the same problem as PApriori and DREBIN that the detection cannot take effect if there are not obvious differences between benign applications and malicious applications for the declared permissions.

Run-Time Performance.
In order to evaluate the applicability of practical implementation of our method, we calculate the time consumption using 50 malicious applications and 50 benign applications, and the size of the applications is from 12 KB to 61 MB.On average, FDP can analyze one application within 15.205 s, which demonstrates that our method is efficient in Android malware detection, though there are many fine-grained information to extract.

Related Work
ere are a lot of research studies related to FDP, which exploits static analysis to extract features such as permissions, API calls, and intent for detection.
According to the differences between malicious apps and benign apps, DroidRanger [36] summarized the rule of permission characteristics to detect unknown malicious apps.Enck et al. developed Kirin [5], which created nine security rules for security.Flet et al. proposed stowaway [37] to detect whether the app declares excessive permission based on the analysis of the mapping relationship between permissions and APIs.DREBIN [7] employed a static method to gather as many features of an application as  possible and use SVM to analyze them.Feizollah et al. [38] proposed the method to detect the malicious applications based on intents and permissions.DroidEnsemble [39] employed string features and structural features for identifying Android malicious applications and optimized the results with ensemble methods.
In this article, FDP differs in three aspects from the previous work.First, we choose the features depending on the characteristic of malicious applications in order to better represent the malicious behavior.Second, the permission features are further subdivided into fine-grained permission features, which enable FDP to detect more malware families, especially, when malicious applications declare few permissions or when the dangerous permissions declared by malicious applications are similar with those declared by benign applications.ird, FDP uses static analysis to extract fine-grained dangerous permission and other features, and it uses the J48 as the optimal classifier to detect malicious applications with TP Rate and F-measure reaching 94.5% and analyze an application in a reasonable time.

Conclusion and Future Work
In this paper, we propose a new method based on the finegrained permissions for detecting Android malicious applications, which gathers the features that better represent the differences between malicious applications and benign applications.e experimental results demonstrate the effectiveness of our FDP method, which shows that FDP can detect more Android malicious families than existed methods.Moreover, the FDP method is efficient enough for practical implementation of Android malware detection.
Although FDP makes a breakthrough compared with the previous work, it still cannot extract more information from dynamic loading, reflection mechanism, and encryption because of inherent limitations of static analysis.Provided that FDP considers these limitations, it must use dynamic analysis method to extract related features, which increases the detection overhead significantly.In addition, more and more android applications try to protect themselves from decompiling, which also increases the difficulty for using our method.erefore, in the premise of efficiency, how to extract the valuable features from dynamic analysis and overcoming the antidisassemble problem is the direction of our future work.Besides, we only use the classic machine learning algorithms in this article, we will try to optimize these machine learning algorithms to construct a better detection model in our future work.Meanwhile, the training data is one important factor to affect the machine learning algorithms to recognize the features pattern.We try to collect more malicious applications and malware families in order to solve the problem of overfitting and underfitting in the future work.

Figure 1 :
Figure 1: Code feature of malicious applications.

Figure 9 :
Figure 9: e top 20 used permissions of 86798 malicious applications.

Table 2 :
Malware families used in the detection model.

Table 3 :
Detection rates with dangerous permissions of features.

Table 4 :
Detection rates with the top 20 permissions of malware samples of features.

Table 5 :
Information gain of dangerous permissions.

Table 6 :
Information gain of the top 20 permissions of 86,798 malware samples.

Table 8 :
Detection rates of the 24 selected permissions.

Table 10 :
Detection rates of FDP.

Table 11 :
e top fourteen features of FDP according to information gain.

Table 12 :
Malware families of the testing set.