Machine Learning-Based Lightweight Android Malware Detection System with Static Features

Abstract

Listen

Translate article

Abstract With the increased popularity and wide adaptation in the embedded system domain, Android has become the main target for malware writers. Due to rapid growth in the number, variants and diversity of malware, fast detection of malware on the Android platform has become a challenging task. Existing malware detection systems built around machine learning algorithms with static, dynamic, or both analysis techniques use a high dimensionality of feature set. Machine learning algorithms are often challenged by such a large number of feature sets and consume a lot of power and time for both training and detection. The model built using such a large number of features may include an irrelevant or negative feature that reduces the overall efficiency of the detection system. In this paper, we present a lightweight Android malware detection system based on machine learning techniques that use fewer static features to distinguish between malicious and benign applications. To reduce the feature dimensions, we have used the feature engineering approach, which utilizes a multilevel feature reduction and elimination process to create a detection model lightweight. Finally, we have built a machine learning-based detection system on the reduced feature set that performs better in comparison to the model build using the original feature set. The proposed detection system achieves accuracy and precision above 98% and drastically reduces the feature set size from 3,73,458 to 105 features.KeywordsAndroidStatic analysisMalwareSecurityMachine learning