On improvements of robustness of obfuscated JavaScript code detection

This paper is dedicated to the problem of design of the detector for obfuscated JavaScript code using machine learning technologies. The main challenge was to design models that would be robust against obfuscators that the model got not familiar with during the training process. During the research we were trying to simulate the scenario when the obfuscation detector, trained to detect samples obfuscated by the specific obfuscators, is given samples that were processed by some another obfuscator. The presented approach of the feature engineering and model training allowed to get better accuracy on the previously unseen obfuscators comparing to the reference work. It was shown that treating minified code samples as obfuscated, as well as enriching the set of the lexical and syntactical features could improve detector’s quality.