Hybrid Phishing Detection Using Stochastic Gradient Descent and Naïve Bayes Optimized with the Mayfly Algorithm

Abstract

Because hackers were able to access AOL user credentials in 1996, phishing, a malicious method of obtaining personal data, became a significant online threat. This fraudulent practice makes use of email and website spoofing techniques to trick victims into disclosing sensitive information. Advanced practices that make use of users' trust and web vulnerabilities, such as spear phishing and tab nabbing, may be hazardous to people's security. In the classification of phishing websites, this research used two prediction models: the Stochastic Gradient Descent (SGD) and the Naïve Bayesian Classification Algorithm (NBC). Hybrid models were developed by incorporating the Mayfly Optimization Algorithm (MOA), a sophisticated optimization method for improving predictive accuracy and overall performance. The dataset contained two stages with a total of 1,353 phishing, trustworthy, and dubious websites. Hyperparameters tuned using random search method for each hybrid model. The dataset contains nine input parameters and derived from previous studies.The results indicated that, with an accuracy of 0.921 during the testing phase, the hybrid model of SGD+MO fared best. On the other hand, the NBC model with Accuracy of 0.877 identified as the weakest model with 4.4% different compared to best model. Also, further improved performance was demonstrated by the numerical classification results for the various categories: it was observed that for phishing websites, the precision metric was 0.925; for suspicious websites, it was 0.933; while for legitimate websites, the precision was 0.911. These results point out the hybrid model's ability to enhance phishing detection systems by showing how well it classifies and detects different kinds of websites.