Man-machine Partial Program Analysis for Malware Detection
With the meteoric rise in popularity of the Android platform, there is an urgent need to combat the accompanying proliferation of malware. Existing work addresses the area of consumer malware detection, but cannot detect novel, sophisticated, domain-specific malware that is targeted specifically at one aspect of an organization (eg. ground operations of the US Military).
Adversaries can exploit domain knowledge to camoflauge malice within the legitimate behaviors of an app and behind a domain-specific trigger, rendering traditional approaches such as signature-matching, machine learning, and dynamic monitoring ineffective. Manual code inspections are also inadequate, scaling poorly and introducing human error. Yet, there is a dire need to detect this kind of malware before it causes catastrophic loss of life and property.
This dissertation presents the Security Toolbox, our novel solution for this challenging new problem posed by DARPA’s Automated Program Analysis for Cybersecurity (APAC) program. We employ a human-in-the-loop approach to amplify the natural intelligence of our analysts. Our automation detects interesting program behaviors and exposes them in an analysis Dashboard, allowing the analyst to brainstorm flaw hypotheses and ask new questions, which in turn can be answered by our automated analysis primitives.
The Security Toolbox is built on top of Atlas, a novel program analysis platform made by EnSoft. Atlas uses a graph-based mathematical abstraction of software to produce a unified property multigraph, exposes a powerful API for writing analyzers using graph traversals, and provides both automated and interactive capabilities to facilitate program comprehension.
The Security Toolbox is also powered by FlowMiner, a novel solution to mine fine-grained, compact data flow summaries of Java libraries. FlowMiner allows the Security Toolbox to complete a scalable and accurate partial program analysis of an application without including all of the libraries that it uses (eg. Android).
This dissertation presents the Security Toolbox, Atlas, and FlowMiner. We provide empirical evidence of the effectiveness of the Security Toolbox for detecting novel, sophisticated, domain-specific Android malware, demonstrating that our approach outperforms other cutting-edge research tools and state-of-the-art commercial programs in both time and accuracy metrics. We also evaluate the effectiveness of Atlas as a program analysis platform and FlowMiner as a library summary tool.
Source: Iowa State University
Author: Thomas Norman Deering