Analysis and Concealment of Malware in an Adversarial Environment

Nowadays, users and devices are rapidly growing, and there is a massive migration of data and infrastructure from physical systems to virtual ones. Moreover, people are always connected and deeply dependent on information and communications. Thanks to the massive growth of Internet of Things applications, this phenomenon also affects everyday objects such as home appliances and vehicles. This extensive interconnection implies a significant rate of potential security threats for systems, devices, and virtual identities. For this reason, malware detection and analysis is one of the most critical security topics. The used detection strategies are well suited to analyze and respond to potential threats, but they are vulnerable and can be bypassed under specific conditions. In light of this scenario, this thesis highlights the existent detection strategies and how it is possible to deceive them using malicious contents concealment strategies, such as code obfuscation and adversarial attacks. Moreover, the ultimate goal is to explore new viable ways to detect and analyze embedded malware and study the feasibility of generating adversarial attacks. In line with these two goals, in this thesis, I present two research contributions. The first one proposes a new viable way to detect and analyze the malicious contents inside Microsoft Office documents (even when concealed). The second one proposes a study about the feasibility of generating Android malicious applications capable of bypassing a real-world detection system. Firstly, I present Oblivion, a static and dynamic system for large-scale analysis of Office documents with embedded (and most of the time concealed) malicious contents. Oblivion performs instrumentation of the code and executes the Office documents in a virtualized environment to de-obfuscate and reconstruct their behavior. In particular, Oblivion can systematically extract embedded PowerShell and non-PowerShell attacks and reconstruct the employed obfuscation strategies. This research work aims to provide a scalable system that allows analysts to go beyond simple malware detection by performing a real, in-depth inspection of macros. To evaluate the system, a large-scale analysis of more than 40,000 Office documents has been performed. The attained results show that Oblivion can efficiently de-obfuscate malicious macro-files by revealing a large corpus of PowerShell and non-PowerShell attacks in a short amount of time. Then, the focus is on presenting an Android adversarial attack framework. This research work aims to understand the feasibility of generating adversarial samples specifically through the injection of Android system API calls only. In particular, the constraints necessary to generate actual adversarial samples are discussed. To evaluate the system, I employ an interpretability technique to assess the impact of specific API calls on the evasion. It is also assessed the vulnerability of the used detection system against mimicry and random noise attacks. Finally, it is proposed a basic implementation to generate concrete and working adversarial samples. The attained results suggest that injecting system API calls could be a viable strategy for attackers to generate concrete adversarial samples. This thesis aims to improve the security landscape in both the research and industrial world by exploring a hot security topic and proposing two novel research works about embedded malware. The main conclusion of this research experience is that systems and devices can be secured with the most robust security processes. At the same time, it is fundamental to improve user awareness and education in detecting and preventing possible attempts of malicious infections.