Malicious Webpage Detection Based on Feature Fusion Using Natural Language Processing and Machine Learning

Malicious websites are purposefully designed to deceive internet users to steal sensitive personal information, infect the victim's system with malware, cause financial losses, and damage the victim's reputation. Finding these pages or links is hard for internet users. Such websites are discovered using detection tools. The majority of detection techniques use blacklisting or whitelisting strategies to find and prevent malicious websites. However, compiling such a sizable list of website links is a time-consuming job that is challenging to update regularly. Therefore, the researchers employ machine learning-based methods to identify these fraudulent connections. These methods are based on the features taken from URLs or web pages. Additionally, features such as DNS details, webpage reputation, and visual similarity data are used. However, these features are few and do not fully utilize the URLs or website contents. This work focuses on merging URL lexical features and content-based features for malicious webpage detection in order to fully exploit the dataset's potential. Natural language processing methods like Hashing, Count, and Term Frequency - Inverse Document Frequency (TF-IDF) vectorizers are employed to extract features from the content of Web pages. The suggested approach's efficiency is evaluated by using the most well-known machine learning methods. The outcome shows that the Count vectorizer with Random Forest achieves a higher accuracy of 91.17% with 500 features.