Date of Award

December 2016

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Heng Yin


Dynamic binary analysis, Dynamic taint analysis, Emulation, Fuzz testing, Security, Virtualization

Subject Categories



Dynamic binary analysis is a prevalent and indispensable technique in program analysis. While several dynamic binary analysis tools and frameworks have been proposed, all suffer from one or more of: prohibitive performance degradation, a semantic gap between the analysis code and the execution under analysis, architecture/OS specificity, being user-mode only, and lacking flexibility and extendability. This dissertation describes the design of the Dynamic Executable Code Analysis Framework (DECAF), a virtual machine-based, multi-target, whole-system dynamic binary analysis framework. In short, DECAF seeks to address the shortcomings of existing whole-system dynamic analysis tools and extend the state of the art by utilizing a combination of novel techniques to provide rich analysis functionality without crippling amounts of execution overhead. DECAF extends the mature QEMU whole-system emulator, a type-2 hypervisor capable of emulating every instruction that executes within a complete guest system environment.

DECAF provides a novel, hardware event-based method of just-in-time virtual machine introspection (VMI) to address the semantic gap problem. It also implements a novel instruction-level taint tracking engine at bitwise level of granularity, ensuring that taint propagation is sound and highly precise throughout the guest environment. A formal analysis of the taint propagation rules is provided to verify that most instructions introduce neither false positives nor false negatives. DECAF’s design also provides a plugin architecture with a simple-to-use, event-driven programming interface that makes it both flexible and extendable for a variety of analysis tasks.

The implementation of DECAF consists of 9550 lines of C++ code and 10270 lines of C code. Its performance is evaluated using CPU2006 SPEC benchmarks, which show an average overhead of 605% for system wide tainting and 12% for VMI. Three platformneutral DECAF plugins - Instruction Tracer, Keylogger Detector, and API Tracer - are described and evaluated in this dissertation to demonstrate the ease of use and effectiveness of DECAF in writing cross-platform and system-wide analysis tools.

This dissertation also presents the Virtual Device Fuzzer (VDF), a scalable fuzz testing framework for discovering bugs within the virtual devices implemented as part of QEMU. Such bugs could be used by malicious software executing within a guest under analysis by DECAF, so the discovery, reproduction, and diagnosis of such bugs helps to protect DECAF against attack while improving QEMU and any analysis platforms built upon QEMU. VDF uses selective instrumentation to perform targeted fuzz testing, which explores only the branches of execution belonging to virtual devices under analysis. By leveraging record and replay of memory-mapped I/O activity, VDF quickly cycles virtual devices through an arbitrarily large number of states without requiring a guest OS to be booted or present. Once a test case is discovered that triggers a bug, VDF reduces the test case to the minimum number of reads/writes required to trigger the bug and generates source code suitable for reproducing the bug during debugging and analysis.

VDF is evaluated by fuzz testing eighteen QEMU virtual devices, generating 1014 crash or hang test cases that reveal bugs in six of the tested devices. Over 80% of the crashes and hangs were discovered within the first day of testing. VDF covered an average of 62.32% of virtual device branches during testing, and the average test case was minimized to a reproduction test case only 18.57% of its original size.


Open Access

Included in

Engineering Commons