Android malware detection using machine learning to mitigate adversarial evasion attacks

Rafiq, Husnain (2022) Android malware detection using machine learning to mitigate adversarial evasion attacks. Doctoral thesis, Northumbria University.

[img]
Preview
Text (Doctoral thesis)
rafiq.husnain_phd(17046430).pdf - Submitted Version

Download (11MB) | Preview

Abstract

In the current digital era, smartphones have become indispensable. Over the past few years, the exponential growth of Android users has made this operating system (OS) a prime target for smartphone malware. Consequently, the arms race between Android security personnel and malware developers seems enduring. Considering Machine Learning (ML) as the core component, various techniques are proposed in the literature to counter Android malware, however, the problem of adversarial evasion attacks on ML-based malware classifiers is understated. MLbased techniques are vulnerable to adversarial evasion attacks. The malware authors constantly try to craft adversarial examples to elude existing malware detection systems. This research presents the fragility of ML-based Android malware classifiers in adversarial environments and proposes novel techniques to counter adversarial evasion attacks on ML based Android malware classifiers.

First, we start our analysis by introducing the problem of Android malware detection in adversarial environments and provide a comprehensive overview of the domain. Second, we highlight the problem of malware clones in popular Android malware repositories. The malware clones in the datasets can potentially lead to biased results and computational overhead. Although many strategies are proposed in the literature to detect repackaged Android malware, these techniques require burdensome code inspection. Consequently, we employ a lightweight and novel strategy based on package names reusing to identify repackaged Android malware and build a clones-free Android malware dataset. Furthermore, we investigate the impact of repacked Android malware on various ML-based classifiers by training them on a clones free training set and testing on a set of benign, non repacked malware and all the malware clones in the dataset. Although trained on a reduced train set, we achieved up to 98.7% F1 score. Third, we propose Cure-Droid, an Android malware classification model trained on hybrid features and optimized using a tree-based pipeline optimization technique (TPoT). Fourth, to present the fragility of Cure- Droid model in adversarial environments, we formulate multiple adversarial evasion attacks to elude the model. Fifth, to counter adversarial evasion attacks on ML-based Android malware detectors, we propose CureDroid*, a novel and adversarially aware Android malware classification model. CureDroid* is based on an ensemble of ML-based models trained on distinct set of features where each model has the individual capability to detect Android malware. The CureDroid* model employs an ensemble of five ML-based models where each model is selected and optimized using TPoT. Our experimental results demonstrate that CureDroid* achieves up to 99.2% accuracy in non-adversarial settings and can detect up to 30 fabricated input features in the best case. Finally, we propose TrickDroid, a novel cumulative adversarial training framework based on Oracle and GAN-based adversarial data. Our experimental results present the efficacy of TrickDroid with up to 99.46% evasion detection.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: mobile systems security, adversarial machine learning, repacked malware detection, adversarial attacks defence, adversarial retraining
Subjects: G400 Computer Science
Department: Faculties > Engineering and Environment > Computer and Information Sciences
University Services > Graduate School > Doctor of Philosophy
Depositing User: John Coen
Date Deposited: 17 Jan 2023 12:16
Last Modified: 17 Jan 2023 12:30
URI: https://nrl.northumbria.ac.uk/id/eprint/51185

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year

View more statistics