COMBINATION OF MULTI-VIEW LEARNING AND DEEP REINFORCEMENT LEARNING TO IMPROVE WEBSITE PHISING DETECTION

Abstract

Phishing is one of the most common and dangerous forms of cyberattacks, where perpetrators attempt to obtain sensitive information by masquerading as trustworthy entities. Traditional detection methods often fail to anticipate new attacks due to the dynamic nature of phishing. This research proposes an adaptive phishing detection system that combines Multi-Kernel Learning (MKL) and Deep Q-Network (DQN) approaches. MKL is utilized to integrate features from URL structure, domain metadata, and webpage content into a rich multi-view representation, while DQN enhances the model's adaptability through a reward-based learning mechanism. This combination was chosen because MKL effectively captures feature variations from different sources, while DQN excels at handling rapidly changing attack patterns. The dataset consists of 11,056 entries with 32 features, divided in an 80:20 ratio for training and testing. Moreover, evaluation is performed using a 5-Fold Cross Validation method to ensure result stability, and hyperparameter exploration is conducted to obtain the best configuration. Evaluation results show that the system achieves an accuracy of 96.34%, precision of 95.8%, recall of 97.85%, F1-score of 96.73%, and AUC of 0.98. These results demonstrate that the MKL-DQN approach is highly effective in accurately and adaptively detecting phishing