Detecting Malicious URLs with Deep Reinforcement Learning

Abstract

One serious problem with network security is the prevalence of online programs that steal data by pretending to be genuine platforms. The vast majority of websites are legitimate, so countermeasures based on artificial intelligence (AI) are often used to identify malicious ones. Here, deep reinforcement learning (DRL) stands up as a promising area for building network intrusion detection models, especially when dealing with very skewed class distributions. But DRL's training time grows exponentially with the complexity of the data. The goal of this research is to expedite training without sacrificing accuracy in classification by integrating a DRL-based classifier with cutting-edge feature selection methods. Five distinct feature-ranking algorithms based on statistics and correlation were utilized in our research on the Mendeley dataset. Compared to the scenario without selection techniques, the findings showed that the selection approach based on the computation of the Gini index greatly improved classification scores, reduced the number of columns in the dataset by 27%, and saved over 10% of training time.