Abstract. Program authorship attribution-identifying a programmer based on stylistic characteristics of code-has practical implications for detecting software theft, digital forensics, and malware analysis. Authorship attribution is challenging in these domains where usually only binary code is available; existing source code-based approaches to attribution have left unclear whether and to what extent programmer style survives the compilation process. Casting authorship attribution as a machine learning problem, we present a novel program representation and techniques that automatically detect the stylistic features of binary code. We apply these techniques to two attribution problems: identifying the precise author of a program, and finding stylistic similarities between programs by unknown authors. Our experiments provide strong evidence that programmer style is preserved in program binaries.

Program binaries are an artifact of a production process that begins with source code and ends with a string of bytes representing executable code. There are many reasons to want to know the specifics of this process for a given binary-for forensic investigation of malware, to diagnose the role of the compiler in crashes or performance problems, or for reverse engineering and decompilation-but binaries are not generally annotated with such provenance details. Intuitively, the binary code should exhibit properties specific to the process that produced it, but it is not at all clear how to find such properties and map them to specific elements of that process.In this paper, we present an automatic technique to recover toolchain provenance: those details, such as the source language and the compiler and compilation options, that define the transformation process through which the binary was produced. We approach provenance recovery as a classification problem, discovering characteristics of binary code that are strongly associated with particular toolchain components and developing models that can infer the likely provenance of program binaries. Our experiments show that toolchain provenance can be recovered with high accuracy, approaching 100% accuracy for some components and yielding good results (90%) even when the binaries emitted by different components appear to be very similar.

We present a novel technique that identifies the source compiler of program binaries, an important element of program provenance. Program provenance answers fundamental questions of malware analysis and software forensics, such as whether programs are generated by similar tool chains; it also can allow development of debugging, performance analysis, and instrumentation tools specific to particular compilers. We formulate compiler identification as a structured learning problem, automatically building models to recognize sequences of binary code generated by particular compilers. We evaluate our techniques on a large set of real-world test binaries, showing that our models identify the source compiler of binary code with over 90% accuracy, even in the presence of interleaved code from multiple compilers. A case study demonstrates the use of inferred compiler provenance to augment stripped binary parsing, reducing parsing errors by 18%.

Binary code presents unique analysis challenges, particularly when debugging information has been stripped from the executable. Among the valuable information lost in stripping are the identities of standard library functions linked into the executable; knowing the identities of such functions can help to optimize automated analysis and is instrumental in understanding program behavior. Library fingerprinting attempts to restore the names of library functions in stripped binaries, using signatures extracted from reference libraries. Existing methods are brittle in the face of variations in the toolchain that produced the reference libraries and do not generalize well to new library versions. We introduce semantic descriptors, high-level representations of library functions that avoid the brittleness of existing approaches. We have extended a tool, unstrip, to apply this technique to fingerprint wrapper functions in the GNU C library. unstrip discovers functions in a stripped binary and outputs a new binary, with meaningful names added to the symbol table. Other tools can leverage these symbols to perform further analysis. We demonstrate that our semantic descriptors generalize well and substantially outperform existing library fingerprinting techniques.

No abstract