Adversarial Attack Defense Research And Implementation Of Malware Detector On Windows Platform

Windows is the most popular desktop operating system in the world,and at the same time,a large amount of malware runs on this platform.In order to curb the spread of malware,machine learning-based malware detection technology has been widely applied.However,the introduction of adversarial machine learning technology has brought new challenges to malware detection.Existing research on adversarial attack defense mainly focuses on image and speech recognition,and the computational cost is high or not applicable to adversarial attack defense in the Windows field.Some researchers have attempted to propose detection methods for adversarial attacks based on PE file endings and segments,but there are still shortcomings such as limited detection of attack types and high false negative rates.In this Paper,we study the adversarial attack defense of the malware detector on the Windows platform,proposes rule-based defense methods and segment boundary recognition-based defense methods,and designs and implements a malware detection system that is immune to adversarial attacks.The main work and achievements are as follows.1.To broaden the detection scope of PE file structure-based adversarial attacks and reduce the false negative rate,we proposed detection methods based on segment boundary recognition and rule matching.Based on the PE file assembly code and data directory entry,we realized the detection of new attacks such as Code Cave through segment boundary recognition,and reduced the false negative rate for Padding,Section Injection,and Filling Slack Space attacks.Based on DOS header,DOS header,and PE header of the PE file,we designed different detection rules to detect new attacks such as Perturb Header Fields,Manipulating DOS Header and Stub,Shift,and Extend.Experimental results show that the proposed methods are superior to other existing methods in terms of attack detection types and accuracy.In addition,experiments also demonstrate that the proposed methods help improve the robustness of commercial malware detectors without modifying them,2,Based on the above adversarial attack defense methods,we designed and implemented an anti-adversarial attack malware detection system.The system provides malware detection functionality for ordinary users.To eliminate the impact of adversarial attacks on malware detection,we detect eight types of structural adversarial attacks before malware detection,restore the perturbed content after detecting adversarial attacks,and perform malware detection on the restored file.In addition,the system provides adversarial attack detection,adversarial attack analysis,and adversarial sample generation functions for malware analysts,as well as system management and maintenance functions for administrators.The system is developed using front-end and back-end separation technology,and various middleware are used to ensure the reliability and ease of scalability of the system.The system uses Redis as a cache database to improve system response speed,and uses RocketMQ to decouple the delivery and execution of detection tasks.At the same time,in order to store a large number of files uploaded by users,MinIO distributed object storage technology is used to improve file access speed and facilitate file management.The system also provides a user-friendly web interface for easy use.