The Technology Research Of Phishing Page Similarity Caculation Based On Visual Features

The phishing attack has been a principal threat to online trading, e-commerce and network information security. It has caused a serious economy and trust crisis to the users and the enterprise, not only jeopardizing the benefit and affecting the life of people, but also influencing the development of e-commerce. There are many problems with websites such as poor design, unregulated source code and templates generated, especially high similarity with official websites. In order to cope with the anti-phishing technique effectively, it is important to design a new scheme to discover new phishing pages.Firstly, we study the computing technology of page similarity from the visual angle, and propose a method to extract visual signature of webpage and a method to compute similarity based on webpage visual signature. We select the pictures and visual blocks of webpage as signature elements, and extract the texture, the location and the text features of signature elements to be visual signature of webpage. We calculate the similarity of visual signature based on EMD algorithm and then combine the feature library to detect whether the webpage is a phishing webpage and to recognize the targets of the phishing webpage.Secondly, we designed a SVM classifier based on location to classify the pictures of the webpage. The accuracy of the classifier is 96.5%. Then we design a method to extract the global texture features of Logo based on subdivided. The sub picture of webpage is seen as the basic unit to describe the webpage, the relationship is very important for estimate the similarity of webpage. We propose a location representation method to reduce the storage space and computing complexity.At last, we design and implement a system based on the similarity of visual features to detect the phishing webpage. Finally, the analysis result of phishing behavior will feed back to the mobile to remind the users. We test this system on a set of phishing instance sample, the accuracy up to 88.64 percent, all of these demonstrate the validity of the system.