Analyzing the Unanalyzable: an Application to Android Apps

In general, software is unreliable. Its behavior can deviate from users’ expectations because of bugs, vulnerabilities, or even malicious code. Manually vetting software is a challenging, tedious, and highly-costly task that does not scale. To alleviate excessive costs and analysts’ burdens, automated static analysis techniques have been proposed by both the research and practitioner communities making static analysis a central topic in software engineering. In the meantime, mobile apps have considerably grown in importance. Today, most humans carry software in their pockets, with the Android operating system leading the market. Millions of apps have been proposed to the public so far, targeting a wide range of activities such as games, health, banking, GPS, etc. Hence, Android apps collect and manipulate a considerable amount of sensitive information, which puts users’ security and privacy at risk. Consequently, it is paramount to ensure that apps distributed through public channels (e.g., the Google Play) are free from malicious code. Hence, the research and practitioner communities have put much effort into devising new automated techniques to vet Android apps against malicious activities over the last decade. Analyzing Android apps is, however, challenging. On the one hand, the Android framework proposes constructs that can be used to evade dynamic analysis by triggering the malicious code only under certain circumstances, e.g., if the device is not an emulator and is currently connected to power. Hence, dynamic analyses can -easily- be fooled by malicious developers by making some code fragments difficult to reach. On the other hand, static analyses are challenged by Android-specific constructs that limit the coverage of off-the-shell static analyzers. The research community has already addressed some of these constructs, including inter-component communication or lifecycle methods. However, other constructs, such as implicit calls (i.e., when the Android framework asynchronously triggers a method in the app code), make some app code fragments unreachable to the static analyzers, while these fragments are executed when the app is run. Altogether, many apps’ code parts are unanalyzable: they are either not reachable by dynamic analyses or not covered by static analyzers. In this manuscript, we describe our contributions to the research effort from two angles: ① statically detecting malicious code that is difficult to access to dynamic analyzers because they are triggered under specific circumstances; and ② statically analyzing code not accessible to existing static analyzers to improve the comprehensiveness of app analyses. More precisely, in Part I, we first present a replication study of a state-of-the-art static logic bomb detector to better show its limitations. We then introduce a novel hybrid approach for detecting suspicious hidden sensitive operations towards triaging logic bombs. We finally detail the construction of a dataset of Android apps automatically infected with logic bombs. In Part II, we present our work to improve the comprehensiveness of Android apps’ static analysis. More specifically, we first show how we contributed to account for atypical inter-component communication in Android apps. Then, we present a novel approach to unify both the bytecode and native in Android apps to account for the multi-language trend in app development. Finally, we present our work to resolve conditional implicit calls in Android apps to improve static and dynamic analyzers.