Enhancing Malware Classification: A Comparison of Filter and Embedded Feature Selection with a Novel Dataset

In recent times, the field of machine learning has witnessed remarkable progress, notably improving the efficiency of malware detection systems. However, the rapid surge in data dimensionality due to advanced technologies necessitates effective feature selection techniques. Feature selection is crucial in refining classifiers by identifying vital features and reducing computational complexity. The multitude of available feature selection algorithms, each with unique criteria, poses a challenge in choosing the right technique for specific datasets in various domains. To tackle this challenge, a combination of Filter and Embedded feature selection methods has been employed. These methods integrates outcomes from multiple feature selection approaches, effectively mitigating the limitations of individual methods. This paper presents a comprehensive comparison between Filter-based techniques, such as chi-squared and Information Gain, ANOVA-F and Embedded techniques like Lasso, Random Forest, XGBoost, and Extra-Tree Classifier. Additionally, it explores API categorization using novel datasets. Experimental findings consistently highlight Random Forest as the preferred choice, consistently delivering high classification accuracy (98%), F-measure (97%), recall (95%), precision (100%), AUC (98%), and demonstrating efficient feature reduction for malware classification datasets. Notably, all feature models exhibit a significant emphasis on Kernel and System Management, Registry Operations, File System and System Information based APIs.