Detection of Malicious HTTP Requests Using Header and URL Features

Cyber attackers leverage the openness of internet traffic to send specially crafted HyperText Transfer Protocol (HTTP) requests and launch sophisticated attacks for a myriad of purposes including disruption of service, illegal financial gain, and alteration or destruction of confidential medical or personal data. Detection of malicious HTTP requests is therefore essential to counter and prevent web attacks. In this work, we collected web traffic data and used HTTP request header features with supervised machine learning techniques to predict whether a message is likely to be malicious or benign. Our analysis was based on two real world datasets: one collected over a period of 42 days from a low interaction honeypot deployed on a Comcast business class network, and the other collected from a university web server for a similar duration. In our analysis, we observed that: (1) benign and malicious requests differ with respect to their header usage, (2) three specific HTTP headers (i.e., accept-encoding, accept-language, and content-type) can be used to efficiently classify a request as benign or malicious with 93.6% accuracy, (3) HTTP request line lengths of benign and malicious requests differ, (4) HTTP request line length can be used to efficiently classify a request as benign or malicious with 96.9% accuracy. This implies we can use a relatively simple predictive model with a fast classification time to efficiently and accurately filter out malicious web traffic.