An Enhanced Anti-Phishing Technique for Social Media Users: A Multilayer Q-Learning Approach

Abstract

As social media usage grows in popularity, so does the risk of encountering malicious Uniform Resource Locator (URLs). Determining the authenticity of a URL can be a highly challenging task, primarily due to the sophisticated attack structure employed by phishing attempts. Phishing exploits the vulnerabilities of computer users, making it difficult to discern between genuine and fraudulent URLs. To address this issue, a self-learning AI framework is required to warn social media users of potentially dangerous links. While several anti-phishing techniques exist, including blacklists, heuristics, and machine learning-based techniques, there is still a need for improvement in terms of detection accuracy. Hence, this study proposed a novel approach to combat phishing attacks using artificial neural networks, and the main aim is to create and validate the anti-phishing technique tool for detection accuracy. Initially, the URL data is collected, followed by preprocessing and then the analysis for malicious activity using the Logistic Bayesian Long Short-Term Memory model (LB-LSTM). The observed malicious URL features are extracted using multilayer Q-learning with the CaspNet and swarm optimization models. Analysis of these features enables the identification of a malicious URL, which is then removed, and the social media user is warned. The proposed technique attained a detection accuracy of 94.33%, Area under the ROC Curve (AUC) of 98.71%, Mean Squared Error (MSE) of 5.67%, Mean average precision of 88.67%, Recall of 98.67%, and F-1 score of 94.34%.