摘要
考虑到在Android终端部署恶意软件检测模型提升性能的同时需要保持准确度,提出一种基于全整型量化与多LSTM的Android恶意软件检测方法。全整型量化通过限制多LSTM网络每层的参数位宽,所有参数将以16位整型的数据类型参与运算,从而缩短检测速度,这些参数也以16位整型存储在存储器中,减少模型内存占用。此外,多LSTM初始网络将对不同语义特征产生不同的权重值,增加语义分布信息,决策网络综合这些权重值进行恶意软件分类,获得较高的准确度。实验结果表明,经过全整型量化后多LSTM检测模型大小缩小到量化前的1/4,检测时间减少约43.9%。与单LSTM相比,基于多LSTM的检测准确度提升了6.7%。
To enhance performance while maintaining accuracy in deploying malware detection models on Android terminals,an Android malware detection method based on full integer quantisation and multi-LSTM is proposed.Full integer quantisation mini-mizes which shortens the speed of detection,these parameters are also stored in memory as 16-bit integer type,reducing the model’s memory footprint.In addition,the initial multi-LSTM network generates different weight values for different semantic features to increase the semantic distribution information,and the decision network synthesizes these weight values for malware classification to obtain higher accuracy.Experiments show that after complete integer quantisation,the size of the multi-LSTM detection model is reduced to 1/4 of its pre-quantisation size,and the detection time is reduced by approximately 43.9%.Com-pared to single LSTM,the accuracy of multi-LSTM based detection shows an improvement of 6.7%.
作者
李红娇
吴佳蓓
顾凡
LI Hong-jiao;WU Jia-bei;GU Fan(College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090 China;COMM.Sector,Nanjing NARI Information and Communication Technology Co.,Ltd,Nanjing 210031,China)
出处
《计算机工程与设计》
北大核心
2025年第8期2280-2288,共9页
Computer Engineering and Design
基金
国家自然科学基金项目(61702321)。
