A Method for Detecting Phishing Websites Based on Tiny-Bert Stacking

Abstract

The Internet is an indispensable part of our lives. Therefore, it is very important to ensure network security and maintain a safe network environment. Phishing, as a low-cost and imperceptible network attack, is rampant in the world’s information networks. To address this issue, this paper proposes a phishing website detection model based on tiny-Bert stacking. The core concept of the proposed model is to use tiny-Bert to extract features from website URL strings, and learn the semantic features and long-range dependent features in URLs. Then we build a Stacking algorithm-based classifier which includes four basic learners among which, CatBoost, XGBoost and LightGBM are the first-level learners, and GBDT is the second-level learner. This detection model can identify phishing websites without manual feature extraction, and the basic learners of Stacking can compensate each other for errors in the classification process, improve generalization, and achieve higher accuracy. The proposed model is evaluated using a dataset based on real phishing websites. Compared to the state of the art, the results show that the proposed model has an accuracy rate of up to 99.14%, a recall rate of up to 99.13%, and is more stable.