The Design And Implementation Of Image Spam Filtering System Based On Cascade Method

In recent years,spammers have embedded spam messages which contained specific promotional purposes into images and sent them to individuals and businesses by emails to avoid the interception of traditional spam filtering systems based on mail texts.This kind of mail is a new type of spam,which is referred to as image-based spam.Compared with text-based spam,imagebased spam usually contains a lot of spam message,such as political and religious propaganda information as well as advertising promotion,which can bring great inconvenience to individuals or business users in their daily work and life.The identification and filtering of image spam has become a hotspot topic for research in the field of information security.therefore,it is of great practical significance to carry out the research on the image spam filtering technology.The traditional text-based spam filtering systems cannot effectively filter the image-based spam,so a co-system is needed as supplementary for existing spam filtering.The research object of this thesis is the image in image-based spam,hereinafter referred to as spam image.Based on the analysis and summarization of the typical characteristics of spam images,this thesis proposes a progressive,two-level filtering system with feedback mechanism.The first layer filter module applied the spam image approximate matching mechanism.The second layer,the convolutional neural network which is a typical model in deep learning is used to classify the spam images.The main work of this thesis is as follows:Firstly,this thesis designed and implemented the module based on approximate matching in the first layer of the filtering system,according to the characteristic of the image spam that it is mostly based on the template production and approximate copy.In this layer,the key factor is to find the suitable image matching algorithm.This thesis designs a series of filtering experiments to test the local feature description algorithm.Based on time and accuracy,the ORB algorithm is used as the approximate matching algorithm in this layer.Secondly,this thesis designed and implemented a deep learning-based filtering module in the second layer of the filtering system based on the strength of the convolutional neural network on the image processing.In this thesis,a concatenated model of convolutional neural network and support vector machine is proposed.The actual spam image database is applied to train the model,the result shows that the classification model can deliver a good classification performance.Finally,this thesis integrated the approximate matching filtering module with the deep learning filtering model into the existing spam filtering system based on the detailed analysis of the overall architecture of the spam filtering system.Based on three comparison experiments,the results show that the cascaded image spam filtering module can effectively improve the processing capacity of the existing spam filtering system for spam images.The cascaded image filtering module can be used as a good supplement to the spam filtering system and the system can play an important role in engineering application.