Automatická detekce malwaru

Abstract

The problem of automatic malware detection presents challenges for antivirus vendors. Since the manual investigation is not possible due to the massive number of samples being submitted every day, automatic malware classication is necessary. Our work is focused on an automatic malware detection framework based on machine learning algorithms. We proposed several static malware detection systems for the Windows operating system to achieve the primary goal of distinguishing between malware and benign software. We also considered the more practical goal of detecting as much malware as possible while maintaining a suciently low false positive rate. We proposed several malware detection systems using various machine learning techniques, such as ensemble classier, recurrent neural network, and distance metric learning. We designed architectures of the proposed detection systems, which are automatic in the sense that extraction of features, preprocessing, training, and evaluating the detection model can be automated. However, antivirus program relies on more complex system that consists of many components where several of them depends on malware analysts and researchers. Malware authors adapt their malicious programs frequently in order to bypass antivirus programs that are regularly updated. Our proposed detection systems are not automatic in the sense that they are not able to automatically adapt to detect the newest malware. However, we can partly solve this problem by running our proposed systems again if the training set contains the newest malware. Our work relied on static analysis only. In this thesis, we discuss advantages and drawbacks in comparison to dynamic analysis. Static analysis still plays an important role, and it is used as one component of a complex detection system.The problem of automatic malware detection presents challenges for antivirus vendors. Since the manual investigation is not possible due to the massive number of samples being submitted every day, automatic malware classication is necessary. Our work is focused on an automatic malware detection framework based on machine learning algorithms. We proposed several static malware detection systems for the Windows operating system to achieve the primary goal of distinguishing between malware and benign software. We also considered the more practical goal of detecting as much malware as possible while maintaining a suciently low false positive rate. We proposed several malware detection systems using various machine learning techniques, such as ensemble classier, recurrent neural network, and distance metric learning. We designed architectures of the proposed detection systems, which are automatic in the sense that extraction of features, preprocessing, training, and evaluating the detection model can be automated. However, antivirus program relies on more complex system that consists of many components where several of them depends on malware analysts and researchers. Malware authors adapt their malicious programs frequently in order to bypass antivirus programs that are regularly updated. Our proposed detection systems are not automatic in the sense that they are not able to automatically adapt to detect the newest malware. However, we can partly solve this problem by running our proposed systems again if the training set contains the newest malware. Our work relied on static analysis only. In this thesis, we discuss advantages and drawbacks in comparison to dynamic analysis. Static analysis still plays an important role, and it is used as one component of a complex detection system.