A machine learning approach to detection of JavaScript-based attacks using AST features and paragraph vectors

Abstract Websites attract millions of visitors due to the convenience of services they offer, which provide for interesting targets for cyber attackers. Most of these websites use JavaScript (JS) to create dynamic content. The exploitation of vulnerabilities in servers, plugins, and other third-party systems enables the insertion of malicious codes into websites. These exploits use methods such as drive-by-downloads, pop up ads, and phishing attacks on news, porn, piracy, torrent or free software websites, among others. Many of the recent cyber-attacks exploit JS vulnerabilities, in some cases employing obfuscation to hide their maliciousness and evade detection. It is, therefore, primal to develop an accurate detection system for malicious JS to protect users from such attacks. This study adopts Abstract Syntax Tree (AST) for code structure representation and a machine learning approach to conduct feature learning called Doc2vec to address this issue. Doc2vec is a neural network model that can learn context information of texts with variable length. This model is a well-suited feature learning method for JS codes, which consist of text content ranging among single line sentences, paragraphs, and full-length documents. Besides, features learned with Doc2Vec are of low dimensions which ensure faster detections. A classifier model judges the maliciousness of a JS code using the learned features. The performance of this approach is evaluated using the D3M dataset (Drive-by-Download Data by Marionette) for malicious JS codes and the JSUNPACK plus Alexa top 100 websites datasets for benign JS codes. We then compare the performance of Doc2Vec on plain JS codes (Plain-JS) and AST form of JS codes (AST-JS) to other feature learning methods. Our experimental results show that the proposed AST features and Doc2Vec for feature learning provide better accuracy and fast classification in malicious JS codes detection compared to conventional approaches and can flag malicious JS codes previously identified as hard-to-detect.