A Privacy-Preserving Phishing Detection System Using Reinforcement Learning and Graph-Based Hybrid Models

Phishing has become a dynamic and frequently emerging attack that threatens the privacy and security of users' data. Traditional phishing detection techniques mostly feature internal privacy constraints, thereby making the new methods inflexible. This article presents a new phishing website detection system known as PhishingGuard+ that integrates Personalized Federated Learning (PFL), Graph Attention Networks (GATs), and Reinforcement Learning (RL). While Reinforcement Learning keeps the system always adaptable to new phishing tactics, Graph Attention Networks create connections between domains and URLs. In addition to that, we propose a Hybrid Model (DCGAN + ResNet50), where ResNet50 improves the accuracy of phish detection, and DCGAN increases the data set. Our approach is based on processing under Personalized Federated Learning, thus working on devices and supporting real-time without compromising the privacy of users. With 99.4% accuracy, 98.7% F1 score, and an AUC of 0.98, PhishingGuard+ beats traditional machine learning systems in various experiments while having flexible, scalable, and privacy-preserving features. Thus, it is this particular study that will allow the future development of adaptive and privacy-friendly systems that can detect phishing techniques.