Research On Adversarial Sample Generation Method For PE Malware Detection

Malware enters people’s field of vision through code implantation,obfuscation,repackaging,etc.Its detection is a hot spot in current research and security product development.,Machine learning is widely employed in the detection field for its advantages of high efficiency and speed,but its efficiency makes people ignore security.Numerous studies have shown that machine learning is susceptible to adversarial attacks,attackers use models to extract features with different emphases then use adversarial techniques to make subtle modifications to malware to conceal the malicious functions of the software.This evolution towards small payloads and high concealment,aimed at misleading the detection model,results in misclassification.Malware is primarily analyzed through its binary code.,the thesis researches malware adversarial samples from the perspective of machine learning security,the main contributions of this thesis are as follows:(1)Aiming at the challenges of malicious function disappearing,file format corruption,and poor execution in the current malware field of adversarial sample generation.To simulate the real-generation environment of samples,this thesis uses black-box attacks.To solve the above problems,this thesis designs two adversarial sample generation models Pa2 AE and ASAE.Additionally,visual technology is employed to evaluate the quality of the generated adversarial samples,to ensure that the scale of the malicious software modifications is minimized.(2)An adversarial sample generation model Pa2 AE based on Pareto multi-objective genetic algorithm is proposed.Firstly,the action strategy library is constructed to realize the modification of malware.Secondly,the fitness objective function is designed according to the problem definition,aiming to maximize the similarity between malware and adversarial samples and minimize the number of modification actions.Combined with the Pareto multiobjective genetic algorithm to evolve the samples,maintain the diversity of samples,and use Cuckoo Sandbox for functional verification during the generation process.Finally,crossvalidation is performed on different attack models and commercial engines,demonstrating the model’s transferability and evasion effectiveness.(3)Based on solving the challenges faced by adversarial samples,in order to reduce the cost waste caused by functional verification and avoid the model vigilance caused by excessive interaction with the classifier,an adversarial sample generation model based on the avoidance sequence adversarial network ASAE is proposed.Firstly,the byte sequence injection algorithm is designed according to the PE file mapping mechanism to ensure sample execution and avoid functional verification.Secondly,the semantic feature information and long dependency of byte sequences are used to construct an evasive sequence adversarial network model for evasion sequence feature learning,the trained model is used for adversarial sample generation.Only the generated sample is used to carry out a single attack on the classifier detection model.Experimental results show that the ASAE model still achieves a high evasion success rate with minimal interactions,demonstrating the model’s stability and transferability.