A Machine Learning Approach for Classifying Java Script Using Static Code Analysis
Abstract
This thesis develops a machine learning approach to classify normal and anomalous JavaScript based on a static analysis of select features derived from the top 30 000 webpages on the internet. A dataset of 136features was extracted from 100 000 raw JavaScript files. Nine test groups were created and tested using 10 subsets of features. K-means clustering was used to group the data and manually translate into binary classification. The results from the K-means clustering show moderate performance with distortions less than 1.0 from elbow plot analysis and average silhouette scores between 0.3 and 0.8 using silhouette analysis of the clustering. The classification of each JavaScript file was then examined using nave Bayes algorithm to re-create and examine the performance of the highest performing classifiers using a less processing intensive method. Nave Bayes was not a good model to re-create the K-means classifier. The best performing classifiers had a Matthews correlation coefficient of 0.75 when examining small JavaScript, and less that 0.38 when examining the medium or large JavaScript. The results show that most JavaScript files were small in file size, and file size was the only defining feature. No features tested effectively categorize the vast majority of JavaScript other than file size. Further research is needed to find features that more accurately encompass the majority of JavaScript to define normal JavaScript.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2022
- Accession Number
- AD1173448
Entities
People
- Michael D. Miller
Organizations
- Naval Postgraduate School