Abstract

Malicious software (Malware) proliferation reaches hundreds of thousands daily. The manual analysis of such a large volume of malware is daunting and time-consuming. The diversity of targeted systems in terms of architecture and platforms compounds the challenges of Android malware detection and malware in general. This highlights the need to design and implement new scalable and robust methods, techniques, and tools to detect Android malware. In this thesis, we develop a malware fingerprinting framework to cover accurate Android malware detection and family attribution. In this context, we emphasize the following: (i) the scalability over a large malware corpus; (ii) the resiliency to common obfuscation techniques; (iii) the portability over different platforms and architectures.

In the context of bulk and offline detection on the laboratory/vendor level: First, we propose an approximate fingerprinting technique for Android packaging that captures the underlying static structure of the Android apps. We also propose a malware clustering framework on top of this fingerprinting technique to perform unsupervised malware detection and grouping by building and partitioning a similarity network of malicious apps. Second, we propose an approximate fingerprinting technique for Android malware's behavior reports generated using dynamic analyses leveraging natural language processing techniques. Based on this fingerprinting technique, we propose a portable malware detection and family threat attribution framework employing supervised machine learning techniques. Third, we design an automatic framework to produce intelligence about the underlying malicious cyber-infrastructures of Android malware. We leverage graph analysis techniques to generate relevant, actionable, and granular intelligence that can be used to identify the threat effects induced by malicious Internet activity associated to Android malicious apps.

In the context of the single app and online detection on the mobile device level, we further propose the following: Fourth, we design a portable and effective Android malware detection system that is suitable for deployment on mobile and resource constrained devices, using machine learning classification on raw method call sequences. Fifth, we elaborate a framework for Android malware detection that is resilient to common code obfuscation techniques and adaptive to operating systems and malware change overtime, using natural language processing and deep learning techniques.

We also evaluate the portability of the proposed techniques and methods beyond Android platform malware, as follows: Sixth, we leverage the previously elaborated techniques to build a framework for cross-platform ransomware fingerprinting relying on raw hybrid features in conjunction with advanced deep learning techniques.