Machine Learning-Based Detection of Smartphone Malware: Challenges and Solutions

The goal of this research is to review the researcher's different attempts with respect to new and emerging technology in malware detection techniques based on machine learning approaches over smartphones. The aim is to evaluate and benchmark these techniques, identify the current landscape of research in this area, and construct a cohesive taxonomy. The available options and gaps will be analyzed to provide valuable insights for researchers regarding the technological environments within this research area. A deep analysis review was conducted to identify studies addressing smartphone security based on machine learning approaches in order to identify all related articles. The outcomes of the last classification scheme of these articles were categorized into types of detection: dynamic analysis, static analysis, hybrid analysis, and uniform resource locator (URL) analysis. The evaluation criteria used in malware detection techniques, with respect to machine learning approaches for smartphones, include accuracy, precision rates (including true positive, false positive, true negative, false negative), training time, f-measure, detection time, area under the curve, true positive, true negative, false positive, false negative, and error rate. Additionally, our classification covers the main machine learning techniques used in the reviewed studies. The taxonomy includes three distinct layers, each reflecting one aspect of the analysis. We also reviewed the details of various types of malicious and benign datasets used within malware detection. Furthermore, open issues and challenges were identified in terms of evaluation and benchmarking, which jeopardize the utilization of this technology. We have described a new recommendation pathway solution that aims to enhance the measurement process of smartphone security applications.