Enhancing Automated Fault Discovery and Analysis

ENHANCING AUTOMATED FAULT DISCOVERY AND ANALYSIS By Jared David DeMott Creating quality software is difficult. Likewise, offensive researchers look to penetrate quality software. Both parties benefit from a scalable bug hunting framework. Once bugs are found, an equally expensive task is debugging. To debug faults, analysts must identify statements involved in the failures and select suspicious code regions that might contain the fault. Traditionally, this tedious task is performed manually. An automated technique to locate the true source of the failure is called fault localization. The thesis of this research is that an automated process to find software bugs and quickly localize the root cause of the failure is possible by improving upon existing techniques. This research is most interested in bugs that lead to security vulnerabilities. These bugs are high value to offensive researchers, and to the typical software test engineer. In particular, memory corruption bugs characterized via an application crash is the subset of all bugs focused on in this work. Existing distributed testing frameworks do not integrate with fault localization tools. Also, existing fault localization tools fail to localize certain difficult bugs. The overall goal of this research is to: (1) Build a dynamic testing framework powerful enough to find new bugs in commercial software. (2) Integrate an existing fault localization technique into the framework that can operate on code without the requirement of having the source code or pre-generated test cases. (3) Create a novel fault localization algorithm that better operates on difficult to localize flaws. (4) Test the improvement on benchmark and real-world code. Those objectives were achieved and empirical studies were conducted to verify the goals of this research. The constructed distributed bug hunting and analysis platform is called ClusterFuzz. The enhanced fault localization process is called Execution Mining. Test results show the novel fault localization algorithm to be an important improvement, and to be more effective than prior approaches. This research also achieved ancillary goals: visualizing fault localization in a new environment; assembly basic blocks for fully compiled code. A pipeline approach to finding and categorizing bugs paves the way for future work in the areas of automated vulnerability discovery, triage, and exploitation.