LGFNN-Stack: A Heterogeneous Stacked Ensemble of LightGBM and Neural Networks for Phishing Detection

Abstract

Phishing attacks are still a major threat to online security, and these types of attacks occur by using deceptive websites that pretend to be legitimate websites and ask for users’ personal information. Various machine learning techniques have been used to detect these types of attacks, and it has been seen that various models have shown limited generalization over different types of data and high false positive rates. This paper proposes a heterogeneous stacked ensemble learning method named LGFNN-Stack, which uses LightGBM and FGNN models. A feedforward neural network (FNN) with a logistic regression meta-model was also developed. Furthermore, a crossdataset feature alignment technique was proposed to enhance the consistency of features as well as prevent information leakage. The proposed model leverages the advantages of the structured learning of the gradient boosting algorithm and the nonlinear representation of neural networks. The experimental results of the cross-dataset validation of the proposed model show that the ROC- AUC value of the hybrid model is 0.992, the precision of the proposed model remains high, and the false positive rate of the proposed model is lower than that of the constituent classifiers. The experimental results prove that the stacked ensemble learning technique can enhance the robustness, reliability, and applicability of the proposed phishing detection system.