| Today,the mobile Internet is highly developed,the online phishing is threatening the privacy and property safety of the people all the time,so it is crucial to detect phishing websites accurately and fast.In this paper,mainstream machine learning algorithms are integrated into a high-performance phishing website detection model by Stacking ensemble learning method,and appropriate improvements are made to it.At the same time,efficient feature engineering is designed and built,including the following aspects:(1)In view of the limitation of most single machine learning algorithms in the field of phishing website detection,a two-level Stacking ensemble learning model is designed to detect phishing websites.Through the early experimental analysis that compares the performance differences of many mainstream machine learning algorithms which are used to build the Stacking model,four most suitable algorithms were selected,and different combinations of Stacking model built by these base learners were compared and analyzed.In the end,XGBoost,Light GBM and DF algorithms were select to build the first level of Stacking model,and GBDT algorithm was used to build the second level.Experimental results show that the Stacking ensemble learning model in this paper has higher accuracy than single mainstream machine learning algorithm.(2)In order to further improve the detection accuracy of Stacking ensemble learning model,the model structure of Stacking algorithm was optimized and improved appropriately in this paper.To making full use of the original features of the data set,the training features of the first level of input Stacking were combined with the detection results of each base learner of the first level as the input features of the second level of Stacking after adding regular penalty terms.At the same time,the detection accuracy of the first-level classification results is evaluated.Each base learner was weighted according to its accuracy rate.The higher the accuracy rate,the higher the weight of the base learner.Then enter the weighted classification results into second level of Stacking model as one of the features.Experimental results show that compared with the general Stacking model,the improved Stacking ensemble learning model in this paper has better detection performance and has significantly improved the detection indicators of phishing websites.(3)In order to dig deeper feature information of phishing websites,this paper designed an HTML string embedding feature based on BERT algorithm,which converted the HTML document of websites into a multidimensional vector through the pre-training process of BERT algorithm.In addition,combined with the URL features of the website and general HTML features,the three kinds of website features were combined as the feature engineering of the phishing website detection model.Through experimental analysis,this feature engineering had improved the accuracy rate of phishing website detection compared with the use of single URL features or general HTML features.Furthermore,the HTML string embedding feature designed based on BERT provided the most effective information in the detection process. |
