神经网络化简非多项式混合布尔算术表达式  

Simplifying Non-polynomial Mixed Boolean-arithmetic Expressions by Neural Network

作　　者：刘彬彬 凤维杰 郑启龙[1,2] 李京[1] LIU Bin-bin;FENG Wei-jie;ZHENG Qi-long;LI Jing(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230026,China;Anhui Key Laboratory of High Performance Computing,University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230026 [2]中国科学技术大学安徽省高性能计算重点实验室,合肥230026

出　　处：《小型微型计算机系统》2023年第3期449-455,共7页Journal of Chinese Computer Systems

基　　金：国家核高基重大专项项目(2012ZX01034-001-001)资助。

摘　　要：混合布尔算术表达式是指混合使用了位运算符和算术运算符的表达式,其是一种先进的软件混淆技术.现有的反混淆方法虽然能够化简特定类型的混合布尔算术表达式,但是对非多项式混合布尔算术表达式仅有有限的化简效果.本文提出一种字符串到字符串的解决方案NeuSim,它通过神经网络来学习和化解非多项式混合布尔算术表达式.首先,本文分别构建基于序列到序列架构和图序列架构的神经网络模型.其次,本文生成一个大规模的非多项式混合布尔算术表达式数据集,它包含一百万个形式多样的表达式样本.在数据集上训练之后,NeuSim可以将一个非多项式混合布尔算术表达式化简为等价的简单表达式.实验结果表明,NeuSim的化简正确率是已有方法的8倍,并且其化简时间低于0.01秒.Mixed Boolean-Arithmetic(MBA)expression, which mixes boolean and arithmetic operations, is an advanced software obfuscation scheme.The existing MBA deobfuscation methods can successfully simplify specific categories of MBA expressions.However, these methods have a limited effect on simplifying non-polynomial MBA expressions.In this paper, we propose a method named NeuSim, a string-to-string solution based on the neural network to learn and simplify non-polynomial MBA expressions.First, we construct sequence-to-sequence and graph-to-sequence neural networks to reduce non-polynomial MBA expressions.Then, we develop a large-scale dataset including one million diversified non-polynomial MBA expressions.After training on the dataset, NeuSim can reduce a non-polynomial MBA expression to an equivalent and concise form.Compared with state-of-the-art work, the evaluation result shows that NeuSim can simplify 8X more non-polynomial MBA expressions, and its simplification time is less than 0.01 seconds.

关 键 词：混合布尔算术表达式 表达式化简 序列到序列神经网络 图序列神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]