A Human-Centric Approach For Binary Code Decompilation

Abstract

Many security techniques have been developed both in academia and industry to analyze source code, including methods to discover bugs, apply taint tracking, or find vulnerabilities. These source-based techniques leverage the wealth of high-level abstractions available in the source code to achieve good precision and efficiency. Unfortunately, these methods cannot be applied directly on binary code which lacks such abstractions. In security, there are many scenarios where analysts only have access to the compiled version of a program. When compiled, all high-level abstractions, such as variables, types, and functions, are removed from the final version of the program that security analysts have access to. <br /> This dissertation investigates novel methods to recover abstractions from binary code. First, a novel pattern-independent control flow structuring algorithm is presented to recover high-level control-flow abstractions from binary code. Unlike existing structural analysis algorithms which produce unstructured code with many goto statements, our algorithm produces fully-structured goto-free decompiled code. We implemented this algorithm in a decompiler called DREAM. Second, we develop three categories of code optimizations in order to simplify the decompiled code and increase readability. These categories are expression simplification, control-flow simplification and semantics-aware naming. We have implemented our usability extensions on top of DREAM and call this extended version DREAM⁺⁺.<br /> We conducted the first user study to evaluate the quality of decompilers for malware analysis. We have chosen malware since it represents one of the most challenging cases for binary code analysis. The study included six reverse engineering tasks of real malware samples that we obtained from independent malware experts. We evaluated three decompilers: the leading industry decompiler Hex-Rays and both versions of our decompiler DREAM and DREAM⁺⁺. The results of our study show that our improved decompiler DREAM⁺⁺ produced significantly more understandable code that outperforms both Hex-Rays and DREAM. Using DREAM⁺⁺participants solved 3 times more tasks than when using Hex-Rays and 2 times more tasks than when using DREAM. Moreover, participants rated DREAM⁺⁺ significantly higher than the competition.