Julia Rubin

The full abstract for this thesis is available in the body of the thesis, and will be available when the embargo expires.

View record

Machine Learning (ML) has been widely applied in different aspects of our lives due to its accuracy and scalability. However, the vulnerability to adversarial examples, which are intentionally designed by attackers to confuse the models, impedes the adoption of them in life- and safety-critical applications. To address this problem, the area of adversarial robustness investigates the mechanisms behind adversarial attacks and defenses against these attacks. Literature in this area exhibits an arm-race trend where defense techniques proposed to address existing attacks were "broken" by newly proposed attacks. A line of research has been conducted in response to such a trend, investigating the reasons for adversarial vulnerabilities, some of them focusing on the inherent limitations of data. Existing surveys on adversarial robustness focus on collecting state-of-the-art attacks and defense techniques, and few of them discuss how the model and/or data explain the adversarial vulnerability observed. In this thesis, we review literature that focuses on the effects of data used to train a model on the model’s adversarial robustness. We systematically identified 57 relevant papers from top publication venues, and categorized them based on the properties of the data discussed. This thesis summarizes the impact of data across eight categories of data properties. Seven of these are general to all applications, and one is specific to a particular application domain. Additionally, we discuss gaps in knowledge and promising future research directions to further improve our understanding of adversarial robustness.

View record

The microservice-based architecture – a SOA-inspired principle of dividing back-end systems into independently deployed components that communicate with each other using language-agnostic APIs – has gained increased popularity in industry. Realistic microservice-based applications contain hundreds of services deployed on a cloud. As cloud providers typically offer a variety of virtual machine (VM) types, each with its own hardware specification and cost, picking a proper cloud configuration for deploying all microservices in a way that satisfies performance targets while minimizing the deployment costs becomes challenging. Existing work focuses on identifying the best VM types for recurrent (mostly high-performance computing) jobs economically. Yet, identifying the best VM type for the myriad of all possible service combinations and further identifying the optimal subset of combinations that minimizes deployment cost is an intractable problem for applications with a large number of services. To address this problem, we propose an approach, called KUBER, which utilizes a set of strategies to efficiently sample the necessary subset of service combinations and VM types to explore. Comparing KUBER with baseline approaches shows that KUBER is able to find the best deployment with the lowest search cost.

View record

This thesis provides a detailed in-depth analysis of Android malware samples that bypassed detection by the Google Play app store and penetrated the official Android market between January 2016 and July 2021. By systematically extracting and analyzing samples from 107 malware families, we identify malicious payloads they execute, conditions guarding execution of the payloads, and other implementation level properties relevant for malware detection. As most samples in our dataset contain multiple payloads, each triggered via its own complex activationlogic, we also contribute a graph-based representation showing multiple activation paths and payloads for each sample in form of a control- and data-flow graph. We discuss the capabilities of existing malware detection tools, put them in context of the properties observed in the analyzed samples, and identify possible gaps and future research directions. We believe that our detailed analysis of the recent, evasive malware families will be of interest to researchers and practitioners and will help further improve malware detection tools. We make our annotated dataset of 1238 samples from 134 malware families available for future studies.

View record

Numerous static taint analysis techniques have recently been proposed for identifying information flows in mobile applications. These techniques are often optimized and evaluated on a set of synthetic benchmarks, which makes the comparison results difficult to generalize. Moreover, the techniques are commonly compared under different configuration setups, rendering the comparisons inaccurate. In this paper, we provide a large, controlled, and independent comparison of the three most prominent static taint analysis tools: FlowDroid, Amandroid, and DroidSafe. We align the configuration setup for the tools and evaluate them on both a set of common benchmarks and on real applications from the Google Play app store. We compare the results of our analysis to the results reported in previous studies, identify main reasons for inaccuracy in existing tools, and provide suggestions for future research.

View record

This thesis proposes a solution for automated goal-driven exploration of Android applications – a scenario in which a user, e.g., security auditor, needs to dynamically trigger the functionality of interest in an application, e.g., to check whether user-sensitive info is only sent to recognized third-party servers. As the auditor might need to check hundreds or even thousands of apps, manually exploring each app to trigger the desired behavior is too time-consuming to be feasible. Existing automated application exploration and testing techniques are of limited help in this scenario as well, as their goal is mostly to identify faults by systematically exploring different app paths, rather than swiftly navigating to the target functionality.The goal-driven application exploration approach proposed in this thesis, called GoalExplorer, automatically generates an executable test script that directly triggers the functionality of interest. The core idea behind GoalExplorer is to first statically model the application UI screens and transitions between these screens, producing a Screen Transition Graph (STG). Then, GoalExplorer uses the STG to guide the dynamic exploration of the application to the particular target of interest: an Android activity, API call, or a program statement. The results of our empirical evaluation on 93 benchmark applications and 95 most popular GooglePlay applications show that the STG is substantially more accurate than other Android UI models and that GoalExplorer is able to trigger a target functionality much faster than existing application exploration.

View record

Microservice-based architecture is a principle inspired by service-oriented approaches for building complex systems as a composition of small, loosely coupled components that communicate with each other using language-agnostic APIs. This architectural principle is now becoming increasingly popular in industry due to its advantages, such as greater software development agility and improved scalability of deployed applications. In this thesis, we report on a broad interview study we conducted, which involved practitioners developing microservice-based applications for commercial use for at least two years. By deliberately excluding “newcomers” and focusing the study on “mature” teams, our goal was to collect best practices, lessons learned, and technical challenges practitioners face. Our study helps inform researchers of challenges in developing microservice-based applications, which can inspire novel software engineering methods and techniques. The study also benefits practitioners who are interested to learn from each other, to borrow successful ideas, and to avoid common mistakes.

View record