Evolutionary Optimization of Neuro-Symbolic Integration for Phishing URL Detection

Abstract

A phishing attack is defined as a type of cybersecurity attack that uses URLs that lead to phishing sites and steals credentials and personal information. Since there is a limitation on traditional deep learning to detect phishing URLs from only the linguistic features of URLs, attempts have been made to detect the misclassified URLs by integrating security expert knowledge with deep learning. In this paper, a genetic algorithm is proposed to find combinatorial optimization of logic programmed constraints and deep learning from given 13 components, which are 12 rule-based symbol components and a neural component. The genetic algorithm explores numerous searching spaces of combinations of 12 rules with deep learning to get an optimal combination of the components. Experiments and 10-fold cross-validation with three different real-world datasets show that the proposed method outperforms the state-of-the-art performance of β -discrepancy integration approach by achieving a 1.47% accuracy and a 2.82% recall improvement. In addition, a post-analysis of the proposed method is performed to justify the feasibility of phishing URL detection via analyzing URLs that are misclassified from either the neural or symbolic networks. © 2021, Springer Nature Switzerland AG.