Survey of Malware Detection through Use of Neural Network Models

In general, antivirus software has heavily relied on signature-based detection methods. These antivirus products usually reference a database of known malicious signatures, making them naturally susceptible to previously unseen malware samples. One of the easiest ways for malware authors to produce these new malware samples is by applying various obfuscation techniques to the previously existing malware. This paper will survey a newer approach to malware detection through neural networks by gathering results from recent journal articles and conference papers published in various online databases. These artificial intelligence-oriented solutions show promise in resilience to unfamiliar malware variants. Explicitly focusing on obfuscated malware, it will compare detection rates between signature-based and neural network-based malware detection systems and organize results using various NN models and training features. Overall, the survey finds that signature-based methods detect between 50 and 80 percent of obfuscated malware samples, while all but one of the neural network methods detect between 90 and 99 percent. This considerable increase in performance suggests that neural networks might provide an effective alternate or supplemental detection method for antivirus products.