An Empirical Evaluation of Machine Learning Algorithms Leveraging Valence-aware Word Embeddings for Web Application Security

With the emergence of Web 2.0, web developers have the ability to generate dynamic web applications, such as online banking, e-commerce, social networking, gaming, and others that revolutionized the industry. However, these web applications have exposed a set of vulnerabilities that can be exploited by malicious users, including broken access controls, cryptographic failures, injection issues, security misconfigurations, and others. Defending against these web application vulnerabilities has become a challenging task. One of the most common ways of detecting and mitigating web vulnerabilities is by utilizing Web Application Firewalls (WAF). WAFs typically use a signature-based technique based on Regular Expressions (RegEx) to create a ruleset to detect multiple known signatures found in malicious HTTP requests. However, recent research has focused on detecting attacks on web applications using machine learning algorithms to classify incoming HTTP requests as malicious or benign. This research leverages a pre-trained word embedding, Valence Aware word embeDding for web Application Security (VADAS), that provides a 1 0-dimensional valence vector that represents each word's maliciousness in the context of web application security. This research performs an empirical evaluation of four machine learning algorithms trained using VADAS valence vectors. The results indicate that the use of valence vectors to classify HTTP requests is an effective method, since each tested algorithm has a precision greater than 98 %. This novel approach provides a new way of detecting web application attacks, relying on minimal interaction from security experts while avoiding the disadvantages of RegEx-based solutions.