• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Evaluation of ensemble learning for Android malware family identification

    2020-04-09 04:01:38JordanWylie譚志遠AhmedAlDubai王建珍
    廣州大學學報(自然科學版) 2020年4期
    關鍵詞:信息

    Jordan Wylie, 譚志遠*, Ahmed Al-Dubai, 王建珍

    (1. School of Computing, Edinburgh Napier University, Edinburgh, EH10 5DT, UK; 2. 山西大學 商務學院/信息學院, 山西 太原 030031)

    Abstract: Each Android malware instance belongs to a specific family that performs a similar set of actions and shares some key characteristics. Being competent to identify Android malware families is critical to address security challenges from malware and the damaging consequences. Ensemble learning is believed an improvement to solve computational intelligence problems by strategically combining decisions from multiple machine learning models. This paper, thus, presents a study of the application of ensemble learning and its effectiveness in Android malware family identification/classification against other single-model-based identification approaches. To conduct a fair evaluation, a prototype of ensemble learning based Android malware classification system was developed for this work, where Weighted Majority Voting (WMV) approach is used in this prototype to determine the importance of individual models (i.e., Support Vector Machine, k-Nearest Neighbour, ExtraTress, Multi-layer Perceptron, and Logistic Regression) to a final decision. The results of the evaluation, conducted by using publicly-accessible malware datasets (i.e., Drebin and UpDroid) and the recent samples from GitHub repositories, show that the ensemble learning approach does not always perform better than single-model learning approaches. The performance of the ensemble learning based malware family classification is heavily influenced by several factors, in particular the features, the values of the parameters and the weights assigned to individual models.

    Key words: Android malware; family identification; static analysis; ensemble learning

    1 Introduction

    Android took the largest market share in the mobile operating system (OS) market in 2012 for the first time ever and which has been continuously growing since then, as shown in Fig.1 that was drawn based on data from[1]. As of May 2019 Android had 2.5 billion monthly active devices[2]. Due to this popularity the threat of malware, which is defined as software that has malicious intent[3-4], increases. Every malware sample belongs to a family of malware that all have a similar set of characteristics, in regards to how they are implemented and their goals. Android malware family identification assists in the identification of steps that could be taken to overturn any damage that has been caused by a specific malware sample[5]. This shows that techniques, being able to identify malware families efficiently and effectively, are vital.

    Fig.1 iOS and Android mobile OS market share[1]

    This paper, therefore, develops an approach to identify Android malware families, in which an ensemble learning algorithm, termed weighted majority voting, is applied to build the malware family classifier. The ensemble learning approach was chosen over other approaches because they perform well when applied to real world issues, improve accuracy, and handle the problems of concept drift and unbalanced data[6]. To fairly evaluate our proposed ensemble learning based malware family identification approach, publicly available evaluation datasets are used. However, the majority of recent related approaches were evaluated on datasets that are out of date which makes it difficult to evaluate how well these related approaches would perform on samples from the current Android malware landscape. This also causes issues when providing comparisons between our proposed approach and these related approaches. To solve these issues, multiple datasets, including commonly-used and recently-discovered malware samples, will be used. Through this research, we aim to answer two main questions defined below.

    (1) How does our ensemble learning approach for Android malware family identification compare to other related single-model-based Android malware classification approaches?

    (2) How effective is our approach developed in classifying recent Android malware into their respective families?

    1.1 Impact

    This work will have an impact on the general public, providing security experts with a weapon to respond to new variants of known malware families effectively which should ultimately result in the competence to define their impact and the steps that could be taken to mitigate any damage caused.

    In addition, the evaluation approach developed in this work will inform how a varied set of classifiers perform while tackling the family identification problem, which will guide the development of new defense approaches.

    1.2 Paper organisation

    The rest of this paper is organised as follows.

    ? Section 2: Provides a relevant technical background, including the structure of an Android application, the different malware analysis evaluation techniques, the evaluation metrics and datasets;

    ? Section 3: Presents a critical review into recent related works;

    ? Section 4: Informed by Section 2, this section proposes a viable ensemble learning approach with an aim to answer the questions defined in Section 1;

    ? Section 5: Critically evaluates the proposed approach making use of multiple metrics discussed in Section 4;

    ? Section 6: Concludes the paper and evaluates how well the questions defined in Section 1 were answered. This section will also provide potential future work.

    2 Technical background

    2.1 Android application structure

    An Android Package (APK) is an archive file that contains various files and directories[7], defined in Table 1.

    Table 1 Contents of an APK file[8]

    There are four different app components that are used by an Android application to provide the system or a user to enter itself. These are activities, services, broadcast receivers, and content providers[7,9]and they are defined in Table 2.

    Table 2 Types of Android components[7,9]

    The activities, services and broadcast receivers are activated by an Intent, which is a messaging object that an Android application use to request an action from another components. A request can be either explicit (where a specific app component will be requested) or implicit (where, in contrast, a type of components will be requested instead). Recent research[9]shows “broadcast receivers” app component was used more often than the other components by malware, and the same components tended to be used across the malware variants of the same family as malware creators reused some of the developed functions.

    To protect user privacy, app permissions are supported by Android, who categorises the permissions into three different groups, namely normal, signature, and dangerous permissions[10]. The 36 normal permissions are low-risk and automatically granted during installation[10-11]. The 29 signature permissions, in contrast, require the requesting and the consuming Android application must be the same[10-11], and two special signature permissions are considered sensitive and should not be used by the majority of applications[10]. The 28 dangerous permissions controlling access to data and private resources, require user confirmation[10].

    2.2 Malware analysis and anti-analysis techniques

    Malware analysis aims to understand the working and intention of malware in order to identify any infected devices and files[12]. This is commonly done through static analysis and dynamic analysis. No execution of the malware of interest is required during the static analysis, whereas the dynamic analysis is in the opposite[12].

    On the other side, to hinder the analysis, anti-analysis techniques have been used in malware creation[13].

    Techniques, such as string encryption and encoding, dynamic loading, native payloads and reflection, help creators to produce evasive malware to bypass dynamic analysis[13-15].

    2.3 Machine learning

    Several metrics have been commonly used by researchers and practitioner to evaluate the effectiveness of trained Machine Learning (ML) models. From the four most basic metrics[16-17], the following advanced metrics shown in Table 3 can be derived.

    Table 3 Main machine learning metrics[18]

    (1)Accuracy: TheAccuracyis the ratio of instances that were classified correctly[16]and is calculated using Equation 1. This metric is effective when classes are balanced. As such, this metric only renders a fair evaluation during the afore-mentioned condition.

    (1)

    (2)Recall: TheRecallshows how many relevant instances were classified into a specific class[11, 16, 17, 19]and is calculated using Equation 2.

    (2)

    (3)Precision: ThePrecisionis the ratio of instances that have been classified correctly against all other instances that were correctly or incorrectly classified to that class[18-19]. This is calculated using Equation 3.

    (3)

    (4)F1-Score: TheF1-Scorecan also be used to evaluate the performance of a model. This is the harmonic mean between the recall and precision metrics[17, 20]and is calculated using Equation 4.

    (4)

    These four advanced metrics will be used in our evaluation.

    2.4 Datasets

    These are the Android Malware Genome Project (AMGP), Drebin, Android Malware Dataset (AMD), UpDroid, Android PRAGuard, AndroZoo, and samples obtained from GitHub.

    (1)AMGP: The AMGP is a dataset that has been used in previous Android malware family identification approaches[5,15], however it has now been discontinued[13, 21]. The dataset contained a total of 1 260 samples that were gathered between August 2010 and October 2011 from multiple marketplaces including the official Google marketplace[22].

    (2)Drebin: The Drebin dataset[23]is another widely used dataset that has been used by various approaches[5, 15]. This dataset has a total of 5 560 samples which were gathered between August 2010 and October 2012 from multiple sources including the Google Play Store, third party marketplaces, and the AMGP[23].

    (3)AMD: The AMD[13]contains 24 560 samples which were obtained from 2010 to 2016 from the Google Play Store, third party security companies, and VirusShare. Each of the samples is associated with a family.

    (4)UpDroid: UpDroid[24]contains 2 479 samples, 50% of which were collected from 2015 onwards. Each of these samples had to be flagged as malicious by at least 20 anti-virus software from VirusTotal. The family name was also derived from the VirusTotal results.

    (5)Android PRAGuard: The Android PRAGuard dataset[14]contains malware from the AMGP and the Contigio dataset. These samples were also obfuscated to form this dataset.

    (6)AndroZoo: The AndroZoo dataset[25]that contains millions of benign and malicious applications. These were obtained through the use of crawlers which were obtained from multiple sources including the Google Play Store, third party marketplaces, torrents, and the AMGP.

    (7)GitHub: There are also multiple repositories on GitHub[26-27]which have more recent samples. The samples obtained from this source would have to be analysed further due to its unreliability.

    Among these seven reviewed datasets, some are more appropriate than the others to be used to evaluate our proposed malware family identification approach. Although the AMGP has been used in several research works. It is outdated and not accessible anymore. The Drebin dataset has also been used to evaluate a significant number of detection approaches, but it is outdated and would not give a real representation of current Android malware too. Similarly, the AMD is an outdated dataset, which however is large in size and could help establishing a better overview of model performance. A more recent dataset. The UpDroid dataset, however, small and has not been used as much. Apart from these old datasets, there are still a few more recent perspective datasets available. The Android PRAGuard dataset is more representative of the current malware landscape as these samples were modified for obfuscation purposes. The AndroZoo dataset could also be used as it is a large dataset in spite of containing benign samples. Last but not least, recent samples from GitHub are accessible but these would not be as reliable as the other datasets mentioned. There are two main criteria to select an evaluation dataset.

    (1)Being able to provide a comparison with other family identification approaches.

    (2)Being able to show how the approach of interest will perform with more recent samples.

    Based on the above criteria, three datasets are selected for this research work. They are the Drebin dataset, UpDroid, and samples obtained from GitHub.

    3 Related works

    Android malware family identification has been widely studied. These related approaches differ themselves from each other by the features used in analysis. The following paragraphs review several key recent related approaches. The quantified evaluation results are summarised in Table 4.

    Table 4 Summary of family identification approaches

    3.1 Approaches using static features

    RevealDroid[5]uses static features in malware family classification. Its main aim was to be resilient to obfuscation techniques. The Classification and Regression Trees (CART) algorithm was applied and a total of 1 054 features were used to build the classifier. This approach was evaluated using a total of 27 979 samples gathered from the AMGP, the Drebin dataset, VirusShare, and VirusTotal. This approach achieved an accuracy of 84%. Whereas no other metrics have been used to evaluate the performance of this approach which makes it difficult to ascertain how well this approach has performed. This is due to the fact that accuracy does not provide a good indication of how well an approach performed when the classes are unbalanced[16, 19].

    Similarly, DroidSieve[15]made use of the ExtraTrees (ET) algorithm and static features in order to infer Android malware families. Four datasets including the AMGP, Drebin dataset, Marvin dataset, and Android PRAGuard were used to evaluate DroidSieve. DroidSieve achieved outstanding performance (i.e., an accuracy of 98.12% and an F-measure of 97.84%) on the Drebin dataset and (i.e., an accuracy of 99.15%) on the AMGP and Android PRAGuard dataset. However, The main issue with this approach is that the Android PRAGuard dataset is a obfuscated version of the AMGP and the Contagio dataset[14]. These obfuscated samples are not samples found in a real life environment.

    3.2 Approaches using dynamic features

    To characterise malware and its runtime behaviour, dynamic features of the malware of interest are extracted from a live environment. A recent research, conducted by Massarelli, et al.[28], made use of Support Vector Machine (SVM) to build a malware classifier with the dynamic features. The Drebin dataset was utilised to evaluate this SVM-based classifier. A70-30 train-test split was applied, and stratified 5-fold cross validation was executed for 20 times. Their approach was able to obtain a mean accuracy of 82%. Although the precision and recall were reported, however, these were calculated for each individual class label (family).

    CANDYMAN[29]involves dynamic features and various classifiers including the Random Forest (RF). This approach was evaluated using the Drebin dataset, from which a total of 4 442 samples from 24 families were taken for evaluation after removing the malware families with fewer than 20 samples. Features of the malware samples were extracted using a Markov chain from the raw features provided through emulation. With the RF classifier, CANDYMAN was able to achieve an accuracy of 81.8%, an f-measure of 80.2%, 80.7% precision and 81.8% recall.

    EnDroid[17]is another approach that made use of dynamic features as well as the stacking classifier. It achieved an accuracy of 94.5% when evaluated against the top 20 malware families with 4 446 samples from the Drebin dataset.

    3.3 Approaches using hybrid features

    Some recent work attempts to take advantages of both static and dynamic features. EC2[30], for example, uses both types of features to train the RF and DBScan classifiers. EC2 is again evaluated using the Drebin dataset, from where the malware samples could not without issues were discarded. The EC2 with the DBScan algorithm achieved a micro f-measure of 76% and a macro f-measure of 97%.

    UpDroid[24]is another hybrid-feature approach equipped with three algorithms, namely k-Nearest Neighbors (kNN), RF, and J48. Alongside with the Drebin dataset, a dataset of 2 479 collected samples was used to evaluate the UpDroid. The samples were executed in an emulated environment for 15 minutes to obtain the dynamic features. Using the newly collected dataset and the kNN algorithm, UpDroid was able to achieve an accuracy of 96.37%, a recall of 96.4% and a false positive rate of 0.2%. On the Drebin dataset, UpDroid achieved an accuracy of 96.85%, a recall of 96.8%, and a false positive rate of 0.3%. This approach has been tested on more recent samples, however the limitation of this approach is the dynamic features used. To extract these dynamic features from the newly collected dataset would take at least 25 days to dynamically analyse every sample.

    Overall, the Drebin dataset was most popular and used to evaluate all the reviewed malware classification approaches. Only UpDroid[24]was evaluated against a more up-to-date dataset. In addition, these approaches were all evaluated with varied metrics which makes it difficult to compare with the approach developed in this paper.

    4 Methodology

    This section defines the process of preparing and organising the datasets, analysing the samples, and defining and evaluating the family identification model. The methodology is summarised in Fig.2.

    Fig.2 Overview of the methodology

    4.1 Dataset selection and preparation

    The datasets were chosen and obtained first. This resulted in a total of 3 datasets being chosen which aimed to answer the research questions stated in Section 1. These datasets are the Drebin dataset[23,31], the UpDroid dataset[24], and samples that were obtained from GitHub[26-27]. These datasets will be split into two experimental datasets. The first experimental dataset contains the samples from the Drebin dataset and will be used to answer our research question 1; the second experimental dataset contains samples from the UpDroid dataset and those collected from GitHub by us, and this second experimental dataset will be used to answer our research question 2.

    To prepare the experimental datasets, all the samples are first organised in a similar manner and then assigned to their respective experimental datasets. Due to the nature of the GitHub samples additional analysis was carried out on those samples. Similar to the sample preparation approach introduced in Ref.[24], only the samples, discovered before May 2019, is kept to ensure they can be correctly recognised by a predefined number (e.g., 20) of major anti-virus scanners. The test results from AVClass, a Python tool to tag malware samples with their respective family label, is then used to build a ground truth of the family labels of the malware samples.

    4.2 Feature extraction

    To extract various static features shown in Table 5 for each sample, AndroGuard, a Python library, is used. The extracted features of the samples are stored in a CSV file. If none of these features can be extracted from a sample, then it is discarded.

    Table 5 Static features used in this project

    4.3 Family identification

    The family identification process steps include data preprocessing, hybrid feature selection, establishment of ensemble learning based malware classification, as well as hybrid-parameter selection. The details of these steps are discussed in the subsections below.

    4.3.1 Data preparation

    The malware family identification begins with data preprocessing. Given a sample setX(defined in Equation 5) that contains s samples with c features, and a respective label setY(defined in Equation 6), the training datasetDis defined as Equation 7.

    Any families having fewer than 5 samples are removed when proceeding with a 5-fold cross validation. The non-Boolean features are then normalised to avoid the ML models from being biased by the features with large values. The Boolean features are not normalised as to keep the property of those features.

    (5)

    (6)

    D=(Y|X)

    (7)

    where:

    s=The number of samples

    c=The number of features

    4.3.2 Hybrid feature selection

    To reduce both the noise in the data and the complexity of computation, feature selection is carried out at this stage. Hybrid feature selection is recommended as exploiting the advantages of both wrapper and filter methods. On one hand, filter methods are not prone to bias and are more time-efficient but do not always improve the effectiveness of a model[32]. On the other hand, wrapper methods usually improve the effectiveness of a model but are time-consuming[32]. As such, it would be wise to first apply a filter method.

    In this work, a correlation-based feature selection, namely Pearson’s correlation as shown in Equation 8[33], is in use. The respective algorithm is defined in Algorithm 1 which removes any features having a correlation greater than 0.8 or less than -0.8. The reason that these values should be chosen is because they indicate a close relationship between the two features[34].

    (8)

    where:

    ρx,y=The Pearson’s correlation between variablesAandB

    μA=The expectation ofA

    μB=The expectation ofB

    σA=The standard deviation ofA

    σB=The standard deviation ofB

    Algorithm 1 Correlation-based feature selection

    Require:Xfeature set,cfeature count

    forx,yinXdo

    CFx,y←ρx,y

    end for

    InitialisecorrF{Empty list that will contain highly correlated features}

    InitialiseremF{Empty list that will contain features to be removed}

    f← list of features

    fori=1 tocdo

    forj=i+1 tocdo

    ifCFi,j<-0.8‖CFi,j>0.8 then

    corrF∪{fi,fj}

    end if

    end for

    end for

    freqs← list of corrF sorted by frequency

    forfeature,freqinfreqsdo

    forCFTincorrFdo

    iffeatureinCFT&& not inremFthen

    remF∪feature

    removeCFTfromcorrF

    end if

    end for

    end for

    Remove features fromXthat are inremF

    returnX

    For the wrapper method, the Recursive Feature Elimination (RFE) algorithm with random forest as an estimator is then used. This works by training a classifier, random forest in this case, to obtain a rank of each feature. Once each feature has a rank, the lowest one is removed, this is repeated until the desired number of features is met[35]. The number is set to 100 in this work. Once the features have been selected, these can then be saved to a new CSV file if required.

    4.3.3 Establishment of malware family identification using ensemble learning with weighted majority vote

    Ensemble learning is used at this stage to build an identifier to classify malware samples. The weighted majority voting mechanism is applied to strategically generate and combine models in the fashion of ensemble learning to solve this classification problem. The ML algorithms used to generate these models are SVM, kNN, ET, Multiplayer Perception (MLP) and Logistic Regression (LR). These ML algorithms have been chosen for their popularity in related research.

    (1)SVM: The SVM classifier is a kernel method algorithm[36]which aims to find the hyperplane that splits the dataset into two classes. Due to this being a multiclass problem a one-vs-rest approach will be taken, where one class is split from the other classes and the class that has the largest distance from the hyperplane is the one that is predicted[36].

    (2)kNN: The kNN classifier predicts the class label by looking at thekclosest samples and is assigned the class label that appears the most[37]. This is known as a lazy algorithm because of the fact that it does not have a training stage[37].

    (3)ET: The ET classifier is similar to the Random Forest classifier used during the wrapper feature selection in that both of these algorithms are ensembles of trees[38]. The difference is that ET makes use of random attributes and cut point values to identify how a node is split[38].

    (4)MLP: The MLP classifier is a feed-forward neural network that uses a set of neurons, in the form of at least one hidden layer which are connected using weights[39]. The weights are altered during training in order to identify the correct weights that will identify the correct class label[39].

    (5)LR: The LR classifier aims to identify the relationship between a feature set and a class label[40]and tends to not over fit as much as other classifiers[41]. Like SVM, LR is a binary classifier but it is possible to use it to handle multiclass problems using a one-vs-rest approach[42].

    To build the ensemble learning based classification approach, the hyperparameters for each of the five chosen classifiers need to be carefully selected. A randomised searching algorithm[43]is applied to pick 10 sets of random parameters and identify the best parameters. The randomised searching algorithm was chosen as its unique properties: (a) a budget can be chosen independent of the number of parameters and possible values; and (b) adding parameters that do not influence the performance does not decrease efficiency.

    The weights for each individual classifier is then calculated using an approach similar to the approach proposed in Ref.[41] where the accuracy of each individual classifier is calculated and divided the sum of the accuracies of the individual classifiers. In this work, the F1-score is used instead due to the fact that accuracy does not perform as well with unbalanced classes[16, 19]. Our improved weighting approach is defined by Equation 9 and Algorithm 2.

    (9)

    where:

    cis the c-th individual classifier

    Wcis the weight of the c-th individual classifier

    F1cis the F1-score of the c-th individual classifier

    Algorithm 2 F1-score based weighting

    Require: A list of classifiers {C1,C2, …,Cn}

    n← the number of classifiers;Wis an empty weight set

    fori=1 tondo

    c←Ci

    W∪{Wc}

    end for

    returnW

    Once the initial candidate set of weights,W, is settled, our proposed malware family identifier based WMV, defined in Equation 10, is evaluated using 5-fold cross validation, where the ensemble model is trained and weights are tuned. Evaluation Metrics, including accuracy, macro F1-score, macro precision, macro recall, and macro AUC, are used alongside a macro ROC curve.

    (10)

    where:

    yis the predicted class of samplex

    mis the number of classifiers

    Cjis the j-th classifier

    wjis the weight of the j-th classifier

    Ais the set of unique class labels

    In order to cater an insight into the types of features that are significant for malware family identification, XGBoost Feature Importance Scores are calculated and the top 20 most important ones are plotted for each experimental dataset. Following the above-introduced process, the experimental results can be reproduced.

    5 Evaluation results and analysis

    Two experiments were conducted and the results are presented in this section. The main goal of the first experiment is to provide a comparison between our ensemble learning based malware family identification approach and other non-ensemble learning based approaches. The second experiment is to evaluate how well our proposed approach performs on recent malware samples.

    5.1 Evaluation results of ensemble learning and its sub-models

    This section delivers the first evaluation of our proposed ensemble learning based malware family identification approach. The performance of the ensemble learning based approach and its sub-model are shown in Table 6. As can be seen from Table 6 the ET and kNN models outperform the proposed ensemble learning based approach, WMV, on the first experimental dataset. This could be due to the limited variation between the weights of the individual classifiers which means that all classifiers were treated equally.

    Table 6 The results of the first experimental dataset %

    Table 7 shows that the proposed ensemble learning based approach, WMV, does outperform the models in terms of F1-score and precision on the second experimental dataset. However, it is not as good as ET with regard to the accuracy and recall. These shortfalls could potentially be improved through further research into weighting calculations.

    Table 7 Experiment 2 results %

    5.2 Comparison between the proposed ensemble learning based malware family identification approach and other approaches

    In order to answer research question 1, a comparison between our ensemble learning based approach (i.e., WMV) and other approaches[15, 17, 24, 28-30]is presented in this section. The different approaches were all evaluated using the same dataset. The evaluation results are shown in Table 8 which details the approaches, the types of features used, and the metrics used to evaluate each approach. These results show that our approach outperforms two out of the three approaches using dynamic features and falls marginally behind the other approaches using static and hybrid features. The reason for this could be due to the reduced feature set. If the feature set was increased and non-linear filter methods were used, there would be a chance that these results could improve at the expense of time. It may also be beneficial to obtain the evaluation scripts for these related approaches. This would provide the ability to evaluate the related approaches using the exact samples as although each of these approaches used the Drebin dataset, certain samples were removed which might impact the results.

    Table 8 Comparison with other approaches that used the Drebin dataset %

    5.3 ROC curves

    As can be seen from Fig.3 the performance of the model using older samples is not perfect which show that only 59% of the time the model would classify a sample correctly. This could be due to various reasons including the fact that the dataset had a total of 83 different labels with only 100 different features to differentiate each of them.

    Fig.3 Experiment 1 macro-average ROC curve

    The performance in experiment 2 was significantly better as can be seen in Fig.4 which shows that 75% of the time the model will classify a sample correctly. This further consolidates the above theory as although the dataset is half the size of experiment 1, it only has a total of 24 families.

    Fig.4 Experiment 2 macro-average ROC curve

    5.4 Feature importance

    Fig.5 and Fig.6 provide the ability to differentiate between which features are important for older samples when compared to newer samples. It can be seen that in general the most important features for more recent samples are increasingly varied as well as some new features (e.g., the signature scheme) added to the list of the top 20.

    Fig.5 Experiment 1 top 20 features

    Fig.6 Experiment 2 top 20 features

    6 Conclusion

    This paper set out to evaluate ensemble learning for malware family classification. To understand the impact of ensemble learning on decision making and its performance in comparison with other non-ensemble learning based malware classification approach, a prototype of ensemble learning-based malware family classification was developed in this paper. The weighted majority voting approach is used in this prototype to determine the importance of individual models to a final decision. The evaluation was conducted using well-known, publicly-available datasets (i.e., Drebin and UpDroid) and the recent malware samples collected from GitHub repositories.

    The evaluation results show that the prototype of ensemble learning based malware family classification outperforms almost all opponents using dynamic features, but marginally behind others using static or hybrid features. The ROC curves show the the prototype has 75% chance being able to distinguish the families of the samples in the second experimental dataset.

    The results imply that the type and quality of features have significant impact on the decision; the randomised searching algorithm does not guarantee optimal selection of parameters. The proposition that ensemble learning outclassed single-model-based learning holds only if the weights of individual models are well tuned. Further investigation is sought to further confirm these implications.

    猜你喜歡
    信息
    訂閱信息
    中華手工(2017年2期)2017-06-06 23:00:31
    展會信息
    中外會展(2014年4期)2014-11-27 07:46:46
    信息超市
    展會信息
    展會信息
    展會信息
    展會信息
    展會信息
    信息
    健康信息
    祝您健康(1987年3期)1987-12-30 09:52:32
    简卡轻食公司| av在线播放精品| 97热精品久久久久久| 99视频精品全部免费 在线| 综合色av麻豆| 精品人妻视频免费看| 久久久久久九九精品二区国产| 国产精品不卡视频一区二区| 波多野结衣巨乳人妻| 亚洲天堂国产精品一区在线| 国产美女午夜福利| 亚洲婷婷狠狠爱综合网| 免费观看人在逋| 亚洲av成人av| 97超视频在线观看视频| 亚洲精品乱久久久久久| 少妇被粗大猛烈的视频| 22中文网久久字幕| 日韩视频在线欧美| 亚洲国产日韩欧美精品在线观看| 建设人人有责人人尽责人人享有的 | 日韩大片免费观看网站 | 国产一区亚洲一区在线观看| 亚洲精品,欧美精品| 久久久a久久爽久久v久久| 日韩一区二区三区影片| 性色avwww在线观看| 一区二区三区四区激情视频| 色吧在线观看| 欧美日本亚洲视频在线播放| 久久精品人妻少妇| 两个人的视频大全免费| 波野结衣二区三区在线| 中国美白少妇内射xxxbb| 亚洲久久久久久中文字幕| 欧美丝袜亚洲另类| 国产黄片视频在线免费观看| 97人妻精品一区二区三区麻豆| 1000部很黄的大片| 日韩av在线免费看完整版不卡| 一区二区三区高清视频在线| 一个人看的www免费观看视频| 亚洲精品乱码久久久v下载方式| 久久久国产成人免费| 非洲黑人性xxxx精品又粗又长| 国产中年淑女户外野战色| 秋霞在线观看毛片| 成人av在线播放网站| 最近手机中文字幕大全| 一区二区三区乱码不卡18| av在线播放精品| 狂野欧美激情性xxxx在线观看| 亚洲精品色激情综合| 高清av免费在线| 久久久成人免费电影| 成人亚洲欧美一区二区av| 欧美区成人在线视频| 九草在线视频观看| 国产伦一二天堂av在线观看| 国产综合懂色| 成人亚洲精品av一区二区| 日本三级黄在线观看| 久久精品国产99精品国产亚洲性色| 最近视频中文字幕2019在线8| 久久这里只有精品中国| 一区二区三区高清视频在线| 在线观看av片永久免费下载| 国产高清不卡午夜福利| 永久免费av网站大全| 久久久亚洲精品成人影院| 亚洲怡红院男人天堂| 久久久久久国产a免费观看| 国产黄片视频在线免费观看| 丰满少妇做爰视频| 干丝袜人妻中文字幕| 欧美97在线视频| 久久精品国产亚洲av涩爱| 日日摸夜夜添夜夜添av毛片| 亚洲国产精品久久男人天堂| 久久久国产成人精品二区| 简卡轻食公司| 国产亚洲一区二区精品| 中文字幕精品亚洲无线码一区| av在线播放精品| 久久久精品94久久精品| 精品午夜福利在线看| 看黄色毛片网站| 午夜福利成人在线免费观看| 能在线免费观看的黄片| 嫩草影院精品99| 美女大奶头视频| 亚洲av日韩在线播放| 观看免费一级毛片| 国产免费又黄又爽又色| 国产乱来视频区| 亚洲国产精品成人久久小说| 亚洲国产精品成人久久小说| 国产成人aa在线观看| 午夜爱爱视频在线播放| 一级黄片播放器| a级毛色黄片| 午夜精品国产一区二区电影 | 禁无遮挡网站| 69av精品久久久久久| 日日干狠狠操夜夜爽| 偷拍熟女少妇极品色| 国国产精品蜜臀av免费| 亚洲自偷自拍三级| 看非洲黑人一级黄片| 亚洲av.av天堂| 草草在线视频免费看| 深夜a级毛片| 久久精品人妻少妇| 成年女人看的毛片在线观看| 国产成人精品久久久久久| av.在线天堂| 亚洲中文字幕一区二区三区有码在线看| 欧美精品一区二区大全| 久久99精品国语久久久| 国产亚洲5aaaaa淫片| 黄片无遮挡物在线观看| 草草在线视频免费看| 成年女人永久免费观看视频| 在线观看美女被高潮喷水网站| 成人午夜高清在线视频| 欧美成人a在线观看| 男人和女人高潮做爰伦理| 99热全是精品| 日本三级黄在线观看| 亚洲,欧美,日韩| 最近中文字幕2019免费版| 婷婷色麻豆天堂久久 | 一边摸一边抽搐一进一小说| 乱系列少妇在线播放| 成人二区视频| 国产亚洲av片在线观看秒播厂 | 好男人视频免费观看在线| 精品久久久久久久末码| 精品久久久久久久末码| videossex国产| 国语对白做爰xxxⅹ性视频网站| 免费观看的影片在线观看| 亚洲欧美日韩高清专用| 久久国内精品自在自线图片| 日韩 亚洲 欧美在线| 久久久久久久国产电影| 亚洲国产精品成人久久小说| 成人午夜高清在线视频| 最近手机中文字幕大全| 极品教师在线视频| 真实男女啪啪啪动态图| 欧美+日韩+精品| 日本黄大片高清| 国产一区二区三区av在线| 国产精品三级大全| 欧美成人一区二区免费高清观看| 色视频www国产| 99久久精品一区二区三区| 天堂中文最新版在线下载 | 男女国产视频网站| 一区二区三区四区激情视频| 国产精品久久久久久av不卡| 日韩欧美精品免费久久| 国产真实乱freesex| 干丝袜人妻中文字幕| 大又大粗又爽又黄少妇毛片口| 国产探花极品一区二区| 国产午夜福利久久久久久| 亚洲av中文av极速乱| 亚洲自拍偷在线| 亚洲美女视频黄频| av又黄又爽大尺度在线免费看 | 性色avwww在线观看| 久久99热6这里只有精品| 国产 一区 欧美 日韩| 看免费成人av毛片| 国产 一区精品| a级毛片免费高清观看在线播放| 亚洲最大成人手机在线| 一级av片app| 一级毛片aaaaaa免费看小| 麻豆成人av视频| av福利片在线观看| 免费看美女性在线毛片视频| 成人漫画全彩无遮挡| 亚洲五月天丁香| 99久久无色码亚洲精品果冻| 大香蕉97超碰在线| 国产视频内射| 青春草视频在线免费观看| 熟妇人妻久久中文字幕3abv| 日本黄色视频三级网站网址| 2021天堂中文幕一二区在线观| 两个人的视频大全免费| ponron亚洲| 成年女人看的毛片在线观看| 村上凉子中文字幕在线| 久久这里只有精品中国| 天堂√8在线中文| 一区二区三区四区激情视频| 一区二区三区乱码不卡18| 免费看av在线观看网站| 国产精品电影一区二区三区| 99久久九九国产精品国产免费| 你懂的网址亚洲精品在线观看 | 爱豆传媒免费全集在线观看| 免费不卡的大黄色大毛片视频在线观看 | 晚上一个人看的免费电影| 日韩欧美国产在线观看| 中文字幕av在线有码专区| 成年版毛片免费区| 国内少妇人妻偷人精品xxx网站| 国产精品av视频在线免费观看| 国产黄色小视频在线观看| 一本久久精品| 国产单亲对白刺激| 亚洲天堂国产精品一区在线| av专区在线播放| kizo精华| 看免费成人av毛片| 永久网站在线| 深夜a级毛片| 亚洲欧洲日产国产| 亚洲欧美清纯卡通| 国产单亲对白刺激| 国产精品野战在线观看| 免费看光身美女| 亚洲av中文av极速乱| 爱豆传媒免费全集在线观看| 国产高清有码在线观看视频| 日韩 亚洲 欧美在线| 十八禁国产超污无遮挡网站| 精品少妇黑人巨大在线播放 | 国产精品三级大全| 岛国在线免费视频观看| 69av精品久久久久久| 久久久国产成人免费| 搡老妇女老女人老熟妇| 中文字幕精品亚洲无线码一区| 国产在视频线在精品| 精品少妇黑人巨大在线播放 | av卡一久久| 久久精品久久久久久久性| 人人妻人人看人人澡| 国内精品一区二区在线观看| 天天躁夜夜躁狠狠久久av| 亚洲伊人久久精品综合 | 国产亚洲av嫩草精品影院| 91精品国产九色| 国产精品野战在线观看| 亚洲欧美日韩东京热| 久久精品91蜜桃| 欧美日韩在线观看h| 69人妻影院| 欧美极品一区二区三区四区| 国产精品国产高清国产av| 午夜福利网站1000一区二区三区| 国产不卡一卡二| 亚洲内射少妇av| 久久6这里有精品| 能在线免费观看的黄片| 亚洲久久久久久中文字幕| 精品久久久久久久人妻蜜臀av| 国产精品1区2区在线观看.| 亚洲精品久久久久久婷婷小说 | 欧美成人a在线观看| 中文字幕制服av| 成年版毛片免费区| 久久久久精品久久久久真实原创| 日本av手机在线免费观看| 免费黄色在线免费观看| 可以在线观看毛片的网站| 你懂的网址亚洲精品在线观看 | 国产单亲对白刺激| 午夜免费男女啪啪视频观看| 午夜福利网站1000一区二区三区| 国产高清视频在线观看网站| 能在线免费看毛片的网站| 欧美3d第一页| 欧美潮喷喷水| 中文亚洲av片在线观看爽| 26uuu在线亚洲综合色| 国产高清国产精品国产三级 | 99久久九九国产精品国产免费| 在线a可以看的网站| 国产一区二区在线观看日韩| 欧美一区二区国产精品久久精品| 国产成人a∨麻豆精品| a级毛色黄片| 国产日韩欧美在线精品| 国产白丝娇喘喷水9色精品| 免费av毛片视频| 搡老妇女老女人老熟妇| 非洲黑人性xxxx精品又粗又长| 日韩一区二区视频免费看| 午夜福利在线观看免费完整高清在| 日本与韩国留学比较| 嫩草影院精品99| 国内精品宾馆在线| 久久人妻av系列| 你懂的网址亚洲精品在线观看 | 成人性生交大片免费视频hd| 国产成人精品一,二区| 人体艺术视频欧美日本| 日韩三级伦理在线观看| 一级毛片我不卡| 中文字幕制服av| 搞女人的毛片| 亚洲精品,欧美精品| 久久精品综合一区二区三区| 午夜福利视频1000在线观看| 欧美极品一区二区三区四区| 看十八女毛片水多多多| 日韩精品有码人妻一区| av国产久精品久网站免费入址| 色尼玛亚洲综合影院| 一级黄色大片毛片| 色噜噜av男人的天堂激情| 成人漫画全彩无遮挡| 成人鲁丝片一二三区免费| 搞女人的毛片| 只有这里有精品99| 国产视频首页在线观看| 国产精品熟女久久久久浪| 亚洲美女搞黄在线观看| 国内精品一区二区在线观看| 2021天堂中文幕一二区在线观| 我要搜黄色片| 国产精品蜜桃在线观看| 岛国毛片在线播放| 国产精品无大码| 嫩草影院新地址| 国产v大片淫在线免费观看| 日本熟妇午夜| 热99在线观看视频| 国产亚洲精品久久久com| 亚洲不卡免费看| 久久精品91蜜桃| 亚洲国产欧美人成| 青春草国产在线视频| 国产成人一区二区在线| 久久这里只有精品中国| videossex国产| 一卡2卡三卡四卡精品乱码亚洲| 亚洲,欧美,日韩| 日韩精品有码人妻一区| 国产免费视频播放在线视频 | 一边摸一边抽搐一进一小说| 天堂影院成人在线观看| 老司机影院毛片| 少妇的逼好多水| 在线免费观看不下载黄p国产| 一级毛片aaaaaa免费看小| 日韩人妻高清精品专区| 97在线视频观看| 少妇人妻精品综合一区二区| 久久久久久久久中文| 91午夜精品亚洲一区二区三区| 色哟哟·www| 久99久视频精品免费| 免费一级毛片在线播放高清视频| 在线观看av片永久免费下载| 国产国拍精品亚洲av在线观看| 日本一二三区视频观看| 非洲黑人性xxxx精品又粗又长| 国产日韩欧美在线精品| 成年女人看的毛片在线观看| 爱豆传媒免费全集在线观看| 秋霞在线观看毛片| 国产激情偷乱视频一区二区| 热99在线观看视频| 青春草国产在线视频| av福利片在线观看| av在线天堂中文字幕| 久久久a久久爽久久v久久| 边亲边吃奶的免费视频| 校园人妻丝袜中文字幕| 人妻系列 视频| 免费看光身美女| 精品国产三级普通话版| 国模一区二区三区四区视频| 久久久a久久爽久久v久久| av在线蜜桃| 国产视频首页在线观看| 男人狂女人下面高潮的视频| 女人久久www免费人成看片 | 亚洲国产精品成人久久小说| 国产在线男女| av在线观看视频网站免费| 我要搜黄色片| 成年av动漫网址| 亚洲电影在线观看av| 国产亚洲最大av| 久久草成人影院| 在线a可以看的网站| 天堂√8在线中文| 久久草成人影院| 插逼视频在线观看| 91久久精品国产一区二区三区| 内地一区二区视频在线| 亚洲av日韩在线播放| 美女xxoo啪啪120秒动态图| 级片在线观看| 久久久午夜欧美精品| 久久精品久久精品一区二区三区| 精品久久久久久久久久久久久| 欧美一区二区精品小视频在线| 国产黄色小视频在线观看| 国产 一区 欧美 日韩| 亚洲精品自拍成人| 好男人视频免费观看在线| 亚洲在线观看片| 国产精品福利在线免费观看| videossex国产| 秋霞伦理黄片| 国内揄拍国产精品人妻在线| 最近手机中文字幕大全| 99久久无色码亚洲精品果冻| 十八禁国产超污无遮挡网站| 亚洲欧美精品专区久久| 亚洲国产色片| 亚洲成人久久爱视频| 久久久久免费精品人妻一区二区| 日本黄色视频三级网站网址| 免费看美女性在线毛片视频| 久久精品夜色国产| 亚洲人成网站在线观看播放| 3wmmmm亚洲av在线观看| 精品一区二区三区视频在线| 黄色配什么色好看| 一级毛片电影观看 | 成人欧美大片| 哪个播放器可以免费观看大片| 成人无遮挡网站| 免费人成在线观看视频色| 国产精品福利在线免费观看| videossex国产| 亚洲av中文字字幕乱码综合| 在线观看一区二区三区| 亚洲四区av| 国产伦在线观看视频一区| 97超视频在线观看视频| 青春草国产在线视频| 三级国产精品欧美在线观看| 久久99热这里只频精品6学生 | 小说图片视频综合网站| 久久久精品欧美日韩精品| 一卡2卡三卡四卡精品乱码亚洲| 白带黄色成豆腐渣| АⅤ资源中文在线天堂| 久热久热在线精品观看| 最近中文字幕2019免费版| 日韩强制内射视频| 内地一区二区视频在线| 午夜日本视频在线| 亚洲五月天丁香| 中国美白少妇内射xxxbb| 日韩大片免费观看网站 | 啦啦啦韩国在线观看视频| 成人三级黄色视频| 一夜夜www| 免费播放大片免费观看视频在线观看 | 国产精品国产三级国产av玫瑰| 日韩 亚洲 欧美在线| 国产白丝娇喘喷水9色精品| 国国产精品蜜臀av免费| 直男gayav资源| 欧美日韩在线观看h| 国产成人aa在线观看| 在线观看66精品国产| 日韩强制内射视频| 又爽又黄a免费视频| 亚洲真实伦在线观看| 最近手机中文字幕大全| av在线亚洲专区| 亚州av有码| 免费看日本二区| 国产高潮美女av| 久久韩国三级中文字幕| 99久久精品一区二区三区| 热99在线观看视频| 在线免费观看不下载黄p国产| 亚洲国产高清在线一区二区三| 国产精品精品国产色婷婷| 真实男女啪啪啪动态图| 美女大奶头视频| 中文乱码字字幕精品一区二区三区 | 久久午夜福利片| 尾随美女入室| 亚洲av福利一区| 成人亚洲精品av一区二区| 91久久精品电影网| 亚洲精品影视一区二区三区av| 亚洲欧美成人精品一区二区| 免费电影在线观看免费观看| 亚洲欧美精品综合久久99| 中文天堂在线官网| 18禁裸乳无遮挡免费网站照片| 国产黄片美女视频| 一个人免费在线观看电影| 波多野结衣巨乳人妻| 亚洲av男天堂| 超碰97精品在线观看| 在线观看av片永久免费下载| 亚洲精品国产av成人精品| 一级黄色大片毛片| 久久99热这里只频精品6学生 | 亚洲成人av在线免费| 综合色丁香网| 亚洲精华国产精华液的使用体验| 又粗又硬又长又爽又黄的视频| 精品99又大又爽又粗少妇毛片| 国产三级在线视频| 能在线免费观看的黄片| 亚洲中文字幕日韩| 亚洲高清免费不卡视频| 国产精品99久久久久久久久| 男女国产视频网站| 美女被艹到高潮喷水动态| 亚洲18禁久久av| 丝袜喷水一区| 亚洲av成人精品一二三区| 亚洲久久久久久中文字幕| 久久国产乱子免费精品| 免费黄色在线免费观看| 国产精品久久久久久久电影| 99久久无色码亚洲精品果冻| 国产午夜精品论理片| 男女边吃奶边做爰视频| av黄色大香蕉| 少妇丰满av| 如何舔出高潮| 久久99蜜桃精品久久| 欧美xxxx性猛交bbbb| 天天躁夜夜躁狠狠久久av| 亚洲成人久久爱视频| 91久久精品电影网| 2021少妇久久久久久久久久久| 全区人妻精品视频| 亚洲va在线va天堂va国产| 日本黄大片高清| 亚洲欧美一区二区三区国产| ponron亚洲| 中文字幕精品亚洲无线码一区| 亚洲av日韩在线播放| 99久久精品国产国产毛片| 中文亚洲av片在线观看爽| 国产在视频线精品| 99久久九九国产精品国产免费| 日韩成人av中文字幕在线观看| 国产精品人妻久久久久久| 日本免费在线观看一区| 99九九线精品视频在线观看视频| 99久久精品一区二区三区| 成人特级av手机在线观看| 国产大屁股一区二区在线视频| 男人舔奶头视频| 久久久色成人| 99视频精品全部免费 在线| 国产黄色小视频在线观看| 亚洲国产日韩欧美精品在线观看| 少妇的逼好多水| 免费观看精品视频网站| 国产av不卡久久| 欧美高清成人免费视频www| 狂野欧美激情性xxxx在线观看| 国产欧美日韩精品一区二区| 亚洲av福利一区| 一级爰片在线观看| 国产亚洲91精品色在线| av在线老鸭窝| 亚洲精品乱久久久久久| av在线老鸭窝| 亚洲av福利一区| 国产成人a区在线观看| 国产一区亚洲一区在线观看| 国产免费又黄又爽又色| 男女边吃奶边做爰视频| 少妇的逼水好多| 国产高清国产精品国产三级 | 亚洲精品456在线播放app| 亚洲欧美精品专区久久| 啦啦啦啦在线视频资源| 亚洲欧美精品专区久久| 久久久国产成人精品二区| 自拍偷自拍亚洲精品老妇| a级毛色黄片| 亚洲乱码一区二区免费版| 亚洲精品456在线播放app| 精华霜和精华液先用哪个| 色视频www国产| 中文欧美无线码| 国产单亲对白刺激| 成人三级黄色视频| 亚洲av不卡在线观看| eeuss影院久久| 亚洲av电影不卡..在线观看| 黄片wwwwww| 亚洲国产精品久久男人天堂| 国产伦精品一区二区三区视频9| 直男gayav资源| 久久久久久伊人网av| 乱码一卡2卡4卡精品| 亚洲av免费在线观看| 国产色婷婷99| 有码 亚洲区| 久久久欧美国产精品| 国产视频首页在线观看| 97超碰精品成人国产| 欧美一区二区亚洲| 久久鲁丝午夜福利片| 国产免费视频播放在线视频 | 日日摸夜夜添夜夜添av毛片| 青青草视频在线视频观看| 伊人久久精品亚洲午夜|