基于特征生成方法的Android恶意软件检测方法
2020年信息技术与网络安全第11期
冯 垚,王金双,张雪涛
陆军工程大学 指挥控制工程学院,江苏 南京210001
摘要: 针对传统特征工程中需要大量专家经验和人力的不足,研究了基于特征生成方法的Android恶意软件检测方法。基于UC Berkeley的ExploreKit自动特征生成方法,通过对原始特征计算获得大量候选特征,根据候选特征的元特征预测其性能并进行评估排序,使用贪心算法从中选出能够提升模型性能的新特征。从APK中提取了敏感API、危险权限等多种特征,在根据信息增益对特征进行筛选后,输入到特征生成框架中,使用C4.5、SVM和随机森林等作为分类模型。实验证明,该方法使错误率平均降低了24.6%,准确率达到了96.5%,曲线下面积(Area Under Curve,AUC)达到了0.99。
中圖分類號: TP393
文獻標識碼: A
DOI: 10.19358/j.issn.2096-5133.2020.11.002
引用格式: 馮垚,王金雙,張雪濤. 基于特征生成方法的Android惡意軟件檢測方法[J].信息技術與網絡安全,2020,39(11):8-13.
文獻標識碼: A
DOI: 10.19358/j.issn.2096-5133.2020.11.002
引用格式: 馮垚,王金雙,張雪濤. 基于特征生成方法的Android惡意軟件檢測方法[J].信息技術與網絡安全,2020,39(11):8-13.
Android malware detection based on feature generation method
Feng Yao,Wang Jinshuang,Zhang Xuetao
Institute of Command Control Engineering,Army Engineering University,Nanjing 210001,China
Abstract: Considering the expert experience and manpower required in traditional feature engineering, a detection method based on feature generation was proposed in this paper. ExploreKit was adopted as the automated feature generation method, which can improve Android malware detection model performance by generating and selecting new features. A large of candidate features are obtaioned by calculating the initial features. According to the meta-features of the candidate features, their performance is predicted and evaluated. New features that will improve the performance of the model are selected by greedy algorithms. The time spent on feature generation is reduced by filtering the initial features input to ExploreKit and limiting the number of generated features. Many features were extracted from APK, such as sensitive API and dangerous permission. By using C4.5, SVM and random forest as the classification models, our experiments show that the error rate of Android malware detection decreases 24.6% on average, the accuracy reaches 96.5%, and the AUC reaches 0.99.
Key words : malware detection;feature engineering;feature generation;ExploreKit
0 引言
機器學習在Android惡意軟件檢測中得到廣泛應用,特征工程是基于機器學習的Android惡意軟件檢測的關鍵環(huán)節(jié)。目前使用的特征主要包括靜態(tài)特征和動態(tài)特征。但特征工程的過程嚴重依賴于專家經驗,反復試驗調優(yōu)才能確定候選特征集合。
針對傳統(tǒng)特征工程需要大量專家經驗和人力的不足,本文提出了基于特征生成方法的Android惡意軟件檢測方法。該方法提取了多類特征,基于UC Berkeley的ExploreKit[1]方法進行自動化特征生成計算,篩選得到能夠提升模型性能的新特征,得到了良好的檢測性能。
本文詳細內容請下載:http://ihrv.cn/resource/share/2000003054
作者信息:
馮 垚,王金雙,張雪濤
(陸軍工程大學 指揮控制工程學院,江蘇 南京210001)
此內容為AET網站原創(chuàng),未經授權禁止轉載。
