Control Flow Obfuscation Using Neural Network To Protect The Logic Information Of Program In Cloud Computing

Cloud computing is an important national information infrastructure,which provides computing platform for Smart City,intelligent manufacturing,financial securities and other industries in the way of service outsourcing.The separation of resource ownership and right of use leads to severe information security problems in cloud computing.By analyzing and tracking the execution process of cloud programs,the attacker or untrusted cloud computing providers can obtain the program execution logic and destroy the confidentiality of cloud program execution.RFCO(runtime control flow obfuscation),a method based on control flow obfuscation,moves the conditions of program branch statements to the private cloud for evaluation,while the interaction between the public cloud and the private cloud generates a lot of time consumption.Aiming at the above problems,this thesis focuses on the protection of program logic based on control flow obfuscation,and proposes an efficient and universal method of program control flow obfuscation based on neural network.The main contributions of this work are as follows:In order to solve the problem of logic leakage in cloud program execution caused by reverse engineering attacks,this thesis proposes a control flow segmentation and transformation method of three address code program based on the soot framework to extract program significant branch statements accurately,and designed the neural network query function to replace the judgement condition of the branch statement.In order to increase the difficulty for malicious attackers to read and understand the program logic,this thesis also proposes a fake-branch statement construction and insertion method based on neural network,which complicates the control flow graph of the program.Furthermore,this thesis proposes a program logic obfuscation framework using neural network,which train neural networks to simulate conditional behaviors of a program.The incomprehensibility of the neural network can effectively protect the program logic.The theoretical analysis shows that the scheme can defend the static reverse analysis,and the single call of the neural network query function is 11.51% of the time cost compared with RFCO solution,which generates a smaller time consumption.Aiming at the passive information acquisition attack through program execution eavesdropping,a loop transformation is proposed to dynamically generate the step value.According to the forward data flow analysis process,this thesis proposes a variable relation mining algorithm,which extracts the constraint relationship between variables before the execution of program statements,conducts data flow consistency audit before the execution of neural network call functions to detect the active attack based on data modification.An efficient execution method based on cache mechanism is proposed to improve the execution efficiency of the obfuscated program computing in cloud.The experiment shows that the cloud efficient execution method reduces the execution time of Monte Carlo algorithm to 42.57% of the original scheme.It is proved by theory and experiment that this scheme can protect the program logic,and the efficiency of obfuscated program execution is high in cloud.Based on the studied techniques and methods,we built a big data cloud computing platform,and a program logic obfuscation system based on neural network was designed and developed to evaluate the reliability and execution efficiency of the scheme.The experimental results show that the control flow obfuscation scheme proposed in this thesis is suitable for big data framework Map Reduce and CPU-sensitive applications.For CPU-sensitive applications,the average time cost of obfuscation scheme is 1.27-2.58 times of the original program.