Phishing attacks continue to pose a significant threat to online security by exploiting user trust through deceptive URLs and malicious web resources. Traditional phishing detection approaches largely rely on isolated feature-based classification techniques, which fail to capture the relational dependencies that naturally exist among web entities such as domains, URLs, and structural attributes. This paper presents a theory-driven framework that models phishing detection as a probabilistic inference problem over a graph structure. URLs and their associated characteristics are represented as interconnected nodes within a graphical model, enabling relational reasoning across the network. To infer the likelihood of phishing behaviour, Loopy Belief Propagation (LBP) is employed as an approximate probabilistic inference mechanism capable of handling cyclic graph structures. The proposed framework emphasizes formal graph construction, probabilistic modelling, and message-passing inference without reliance on implementation-specific heuristics. By reasoning collectively over relational dependencies, the model provides a robust theoretical foundation for phishing detection under uncertainty. This work contributes a formalized approach that bridges graph- based learning and probabilistic inference for cybersecurity applications.
Hole et al. (Thu,) studied this question.