The detection of similar code can support many so ware engineering tasks such as program understanding and program classification. Many excellent approaches have been proposed to detect programs having similar syntactic features. However, these approaches are unable to identify programs dynamically or statistically close to each other, which we call behaviorally similar programs. We believe the detection of behaviorally similar programs can enhance or even automate the tasks relevant to program classification. In this thesis, we will discuss our current approaches to identify programs having similar behavioral features in multiple perspectives.
We first discuss how to detect programs having similar functionality. While the definition of a program’s functionality is undecidable, we use inputs and outputs (I/Os) of programs as the proxy of their functionality. We then use I/Os of programs as a behavioral feature to detect which programs are functionally similar: two programs are functionally similar if they share similar inputs and outputs. This approach has been studied and developed in the C language to detect functionally equivalent programs having equivalent I/Os. Nevertheless, some natural problems in Object Oriented languages, such as input generation and comparisons between application-specific data types, hinder the development of this approach. We propose a new technique, in-vivo detection, which uses existing and meaningful inputs to drive applications systematically and then applies a novel similarity model considering both inputs and outputs of programs, to detect functionally similar programs. We develop the tool, HitoshiIO, based on our in-vivo detection. In the subjects that we study, HitoshiIO correctly detect 68.4% of functionally similar programs, where its false positive rate is only 16.6%.
In addition to functional I/Os of programs, we attempt to discover programs having similar execution behavior. Again, the execution behavior of a program can be undecidable, so we use instructions executed at run-time as a behavioral feature of a program. We create DyCLINK, which observes program executions and encodes them in dynamic instruction graphs. A vertex in a dynamic instruction graph is an instruction and an edge is a type of dependency between two instructions. The problem to detect which programs have similar executions can then be reduced to a problem of solving inexact graph isomorphism. We propose a link analysis based algorithm, LinkSub, which vectorizes each dynamic instruction graph by the importance of every instruction, to solve this graph isomorphism problem efficiently. In a K Nearest Neighbor (KNN) based program classification experiment, DyCLINK achieves 90 + % precision.
Because HitoshiIO and DyCLINK both rely on dynamic analysis to expose program behavior, they have better capability to locate and search for behaviorally similar programs than traditional static analysis tools. However, they suffer from some common problems of dynamic analysis, such as input generation and run-time overhead. These problems may make our approaches challenging to scale. Thus, we create the system, Macneto, which integrates static analysis with machine topic modeling and deep learning to approximate program behaviors from their binaries without truly executing programs. In our deobfuscation experiments considering two commercial obfuscators that alter lexical information and syntax in programs, Macneto achieves 90 + % precision, where the groundtruth is that the behavior of a program before and after obfuscation should be the same.
In this thesis, we offer a more extensive view of similar programs than the traditional definitions. While the traditional definitions of similar programs mostly use static features, such as syntax and lexical information, we propose to leverage the power of dynamic analysis and machine learning models to trace/collect behavioral features of pro- grams. These behavioral features of programs can then apply to detect behaviorally similar programs. We believe the techniques we invented in this thesis to detect behaviorally similar programs can improve the development of software engineering and security applications, such as code search and deobfuscation.
Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8MW3V2Z |
Date | January 2018 |
Creators | Su, Fang-Hsiang |
Source Sets | Columbia University |
Language | English |
Detected Language | English |
Type | Theses |
Page generated in 0.002 seconds