DETECTION OF ANDROID MALWARE USING DEEP LEARNING ENSEMBLE WITH CHEETAH-OPTIMIZED FEATURE SELECTION

Abstract

On account of the rapid growth with mobile threats, it is necessary to safeguard the user security and privacy. Our goal in this article is to develop robust algorithms capable of distinguishing accurately between benign and malicious apps. The paper provides a methodology for Android malware detection, integrating preprocessing, extraction and selection of feature and a novel deep learning-based detection model. The aim of this approach is to address the challenges that are associated with the imbalanced datasets, ensuring a balanced representation of benign and samples through the effective data collection. The Synthetic Minority Oversampling Technique (SMOTE) has been employed for oversampling the minority class that enhances the dataset balance, for imbalance cases. The extraction of feature involves the statistical feature extraction, which includes the measures like standard deviation, median, mean, variance, interquartile range (IQR), skewness, and kurtosis. Flow-based features such as Flow ID and Flow Duration are additionally considered to capture the essential characteristics of the data. For further refining the feature set, a combination of cheetah optimization algorithm (COA) and the variable velocity strategy particle swarm optimization algorithm (VVS-PSO) is employed in selection of features. This hybrid approach is enhanced by the discriminatory power of selected features, which improves the effectiveness of the model. The proposed deep learning-based detection model is named DSDNN (Deep Siamese+DenseNet+NasNet). Ensemble combined with the deep Siamese network for similarity measurement, DenseNet is used for efficient feature extraction, and NasNet is used to enhance learning capabilities. The ensemble of these classifiers ensures a robust and versatile model for the detection of Android malware.