Malicious Urls Detection Using Deep Learning

In recent years,with the rapid development of the Internet,websites registering becomes extremely easy.So,the number and size of websites grow rapidly.On one hand,this has become an important part of our everyday life for information communication and knowledge dissemination.It helps to transact information timely,rapidly and easily.On the other hand,Malicious URLs host unsolicited content(spam,phishing,drive-by exploits,etc.)and lure unsuspecting users to become victims of scams(monetary loss,theft of private information,and malware installation),and cause losses of billions of dollars every year.it also provides living space for malicious websites.URL detection is one of the most important technologies to prevent attacks from malicious websites.However,to detect such crimes systems should be fast and precise with the ability to detect new malicious content.Traditionally,this detection is done mostly through the usage of blacklists.However,blacklists cannot be exhaustive,and lack the ability to detect newly generated malicious URLs.As malicious URL becomes more complex,traditional URL detection methods seem to be unable to handle it.Thus,it is particularly important to explore new,more reliable and accurate URL detection methods.Our work is a high-precision,susceptible-learning and rapidly-reacting malicious URL detection system based on the deep learning algorithms.The whole work can be divided into three parts:the features selecting and processing,the training part and the classification part.In the features selecting and processing,we get the word2vec feature,TF-IDF feature and content-based feature respectively.These features have been used to train different machine learning and deep learning algorithms to detect and predicting URLs it is good or bad.In the results section,the performance evaluation of each of these algorithms have been presented.The results show that the algorithns Convolutional Neural Network(CNN)achieved the highest accuracy to compared with two other algorithms to know(SVM,LR).In the final classification stage,we use a layer of a fully connected network to further address the three probability matrices,and give a final result(malicious,benign).