An Assessment of Lexical, Network, and Content-Based Features for Detecting Malicious URLs Using Machine Learning and...

Area of Science:

• Cybersecurity
• Machine Learning
• Web Security

Background:

• World Wide Web services are integral to daily life, accessed via Uniform Resource Locators (URLs).
• Malicious actors exploit URLs to create deceptive websites, leading to spam, spyware, phishing, and malware attacks.
• Effective detection of malicious URLs is critical for preventing cybercriminal activities.

Purpose of the Study:

• To evaluate machine learning (ML) and deep learning (DL) models for malicious URL detection.
• To engineer and analyze lexical-based, network-based, and content-based features for improved detection.
• To identify the most effective features and models for predicting malicious URLs.

Main Methods:

• Utilized a dataset of 66,506 URL records.
• Engineered lexical, network, and content-based features.
• Applied feature selection algorithms: correlation analysis, ANOVA, and chi-square.
• Compared the performance of various ML and DL models.

Main Results:

• Naïve Bayes (NB) demonstrated superior performance among the evaluated models.
• The Naïve Bayes model achieved an accuracy of 96% in detecting malicious URLs.
• Feature engineering and analysis identified key discriminative features for prediction.

Conclusions:

• The study successfully identified effective features and models for malicious URL detection.
• Naïve Bayes proved to be the most accurate model for this task within the study's scope.
• The research contributes a robust methodology and findings for enhancing web security against malicious URLs.