基于改进的MobileNetV2模型的安卓恶意家族分类方法研究

Research on Classification Method of Android Malware Family based on Improved MobileNetV2 Model

针对人类视觉系统对颜色的高度敏感特点,提出一种基于改进的MobileNetV2模型的安卓恶意家族分类方法。该方法通过引入注意力机制,对RGB图像的3个通道进行特征融合,提高模型对图像颜色信息的敏感度。同时,针对小样本数据集的问题,提出一种改进的模块结构,减少模型的深度和宽度,以提高模型对小样本数据集的特征提取能力。将SE(squeeze-and-excitation network)注意力机制与CBAM(convolution block attention module)注意力机制融入模型进行对比,实验结果表明:CBAM注意力机制在该图像分类任务中表现出显著的优越性,准确率达到94.18%,比原有模型提高了3.16%,验证了该方法的有效性和实用性。该研究对于小样本数据集的图像分类任务的准确性和实际应用中的性能有重要意义。

Aiming at the highly sensitive characteristics of the human visual system to color,a classification method of the Android malicious family based on the improved MobileNetV2 model is proposed.By introducing the attention mechanism,this method performs feature fusion on the three channels of RGB image to improve the sensitivity of the model to the color information of the image.At the same time,aiming at the problem of small-sample datasets,an improved module structure is proposed,which reduces the depth and width of the model and improves the feature extraction ability of the model for small-sample datasets.The experimental results show that the Sequeeze-and-Excitation Network(SE)and the Convolution Block Attention Module(CBAM)are both located in the model.The experimental results show,the CBAM attention mechanism shows significant superiority in this image classification task,with an accuracy rate of94.18%,which is 3.16%higher than the original model,which verifies the effectiveness and practicality of the proposed method.This study has important implications for the accuracy of image classification tasks and the performance in practical applications of small-sample datasets.