基于卷积神经网络恶意安卓应用行为模式挖掘

Mining behavior pattern of mobile malware with convolutional neural network

现有的安卓恶意应用检测方法所提取的特征冗余且抽象,无法在高级语义上反映恶意应用的行为模式。针对这一问题,提出一种可解释性检测方法,通过社交网络检测算法聚类可疑系统调用组合,将其映射为单通道图像,用卷积神经网络进行分类,并利用卷积层梯度权重类激活映射可视化方法发现最可疑的系统调用组合,从而挖掘理解恶意应用行为。实验结果表明,所提方法在高效检测的基础上,能够正确发现恶意应用的行为模式。

The features extracted by existing malicious Android application detection methods are redundant and too abstract to reflect the behavior patterns of malicious applications in high-level semantics. In order to solve this problem, an interpretable detection method was proposed. Suspicious system call combinations clustering by social network analysis was converted to a single channel image. Convolution neural network was applied to classify Android application. The model trained was used to find the most suspicious system call combinations by convolution layer gradient weight classification activation mapping algorithm, thus mining and understanding malicious application behavior. The experimental results show that the method can correctly discover the behavior patterns of malicious applications on the basis of efficient detection.