MA_BiRAE - Malware analysis and detection technique using adversarial learning and deep learning

Abstract

Malware attacks are frequently increasing due to the growing use of handheld gadgets, especially Android phones. Hackers try to access smartphones through a variety of techniques, including the theft of information, tracking, and deceptive advertising. There are various techniques for malware analysis and detection, but some issues, like low performance, computational complexity, overfitting, and so on, have been identified while detecting malware and training data. To address these issues, the proposed technique is designed to achieve efficient malware detection. Initially, data is collected from the Aposemat IoT-23 and Bot-IoT datasets, and the Adaptative Perturbation Pattern Method (Ap2 m) is used to generate constrained adversarial samples. Evasion attacks are used to examine regular adversarial training, while Improved Random Forest (IRF) is used for modeling and fine-tuning. The deep Residual Convolutional Neural Network (deep RCNet) is utilized to extract the features. Finally, the Multi-head Attention-based Bidirectional Residual Autoencoder (MA_BiRAE) model is used for malware detection. The performance of the proposed technique is compared to various existing models to determine its superiority. The proposed technique is evaluated using two datasets: the Aposemat IoT-23 dataset and the Bot-IoT dataset. The proposed technique achieves an accuracy of 99.63% for the Aposemat IoT-23 dataset and 99.11% for the Bot-IoT dataset.