WYSINWYX: What You See Is Not What You eXecute

Gogul Balakrishnan
University of Wisconsin

There is an increasing need for tools to help programmers and security analysts understand executables. For instance, commercial companies and the military increasingly use Commercial Off-The Shelf (COTS) components to reduce the cost of software development. They are interested in ensuring that COTS components do not perform malicious actions (or can be forced to perform malicious actions). Viruses and worms have become ubiquitous. A tool that aids in understanding their behavior can ensure early dissemination of signatures, and thereby control the extent of damage caused by them. In both domains, the questions that need to be answered cannot be answered perfectly -- the problems are undecidable -- but static analysis provides a way to answer them conservatively.

In recent years, there has been a considerable amount of research activity to develop analysis tools to find bugs and security vulnerabilities. However, most of the effort has been on analysis of source code, and the issue of analyzing executables has largely been ignored. In the security context, this is particularly unfortunate, because performing analysis on the source code can fail to detect certain vulnerabilities due to the WYSINWYX phenomenon: ``What You See Is Not What You eXecute''. That is, there can be a mismatch between what a programmer intends and what is actually executed on the processor.

Even though the advantages of analyzing executables are appreciated and well-understood, there is a dearth of tools that work on executables directly. The overall goal of our work is to develop algorithms for analyzing executables, and to explore their applications in the context of program understanding and automated bug hunting. Unlike existing tools, we want to provide useful information about memory accesses, even in the absence of debugging information. Specifically, the dissertation focuses on the following aspects of the problem:

The algorithms described in this dissertation are incorporated in a tool we built for analyzing Intel x86 executables, called CodeSurfer/x86.

Because executables do not have a notion of variables similar to the variables in programs for which source code is available, one of the important aspects of IR recovery is to determine a collection of variable-like entities for the executable. The quality of the recovered variables affects the precision of an analysis that gathers information about memory accesses in an executable, and therefore, it is desirable to recover a set of variables that closely approximate the variables of the original source-code program. On average, our technique is successful in identifying correctly over 88% of the local variables and over 89% of the fields of heap-allocated objects. In contrast, previous techniques, such as the one used in the IDAPro disassembler, recovered 83% of the local variables, but 0% of the fields of heap-allocated objects.

Recovering useful information about heap-allocated storage is another challenging aspect of IR recovery. We propose an abstraction of heap-allocated storage called recency-abstraction, which is somewhere in the middle between the extremes of one summary node per malloc site and complex shape abstractions. We used the recency-abstraction to resolve virtual-function calls in executables obtained by compiling C++ programs. The recency-abstraction enabled our tool to discover the address of the virtual-function table to which the virtual-function field of a C++ object is initialized in a substantial number of cases. Using this information, we were able to resolve, on average, 60% of the virtual-function call sites in executables that were obtained by compiling C++ programs.

To assess the usefulness of the recovered IR in the context of bug hunting, we used CodeSurfer/x86 to analyze device-driver executables without the benefit of either source code or symbol-table/debugging information. We were able to find known bugs (that had been discovered by source-code analysis tools), along with useful error traces, while having a low false-positive rate.

(Click here to access the dissertation: PDF.)