Building a Machine Learning Classifier for Malware Detection

Current signature-based antivirus software is ineffective against many modern malicious software threats. Machine learning methods can be used to create more effective antimalware software, capable of detecting even zero-day attacks. Some studies have investigated the plausibility of applying machine learning to malware detection, primarily using features from n-grams of an executables file's byte code. We propose an approach that primarily learns from metadata, mostly contained in the headers of executable files, specifically the Windows Portable Executable 32-bit (PE32) file format. Our experiments indicate that executable file metadata is highly discriminative between malware and benign software. We also employ various machine learning methods, finding that Decision Tree classifiers outperform Logistic Regression and Naive Bayes in this setting. We analyze various features of the PE32 header and identify those most suitable for machine learning classifiers. Finally, we evaluate changes in classifier performance when the malware prevalence (fraction of malware versus benign software) is varied.