Component features based enhanced phishing website detection system using EfficientNet, FH-BERT, and SELU-CRNN methods

Abstract

Introduction: Phishing is a type of cybercrime used by hackers to steal sensitive user information, making it essential to detect phishing attacks on websites. Many prevailing works have utilized Uniform Resource Locator (URL) links and Document Object Model (DOM) tree structures for Phishing Website Detection (PWD). However, since phishing websites imitate legitimate websites, these approaches often produce inaccurate detection results. Methods: To enhance detection efficiency, we propose a PWD system that focuses on important website features and components. The process begins with collecting URL links from phishing website datasets, followed by the generation of Hypertext Markup Language (HTML) formats. A DOM tree structure is then constructed from the HTML, and components are extracted along with Natural Language Processing (NLP) features, credentials, URL, DOM tree similarity, and component features. The DOM-tree components are converted into score values using Feature Hasher-Bidirectional Encoder Representations from Transformers (FH-BERT). These score values are fused with component features, and significant features are selected using an Entropy-based Chameleon Swarm Algorithm (ECSA). Results: The final classification is performed by Scaled Exponential Linear Unit Convolutional Recurrent Neural Network (SELU-CRNN). Simulation results demonstrate that the proposed technique improves PWD performance, achieving higher accuracy (98.42%) and reduced training time (63,003 ms) compared to prevailing methods. Discussion: By integrating component, semantic, and structural features, the proposed model enhances both robustness and efficiency, making it an effective solution for phishing website detection.