Ensemble of Multi-Spatial Resolution for Image Spam Filtering

Abstract

Image spam is a type of spam e-mail that contains an image in the body of the e-mail and holds malware and other malicious threats. The rise of image spam has become a serious concern for e-mail users. This paper presents a spam image classification scheme with two primary goals. Firstly, Multi Spatial Resolution (MSR) with four different levels of resolution is proposed to improve the representation of images by incorporating spatial information between features. Due to the fact that MSR generates distinct image representations for each level, the predictions obtained from each representation may give different results. Thus, the final prediction of whether an image is spam or legitimate is difficult to determine. To solve this problem, an ensemble of MSR is proposed to combine the class probabilities of the model at each level to obtain a final prediction. The experiment was carried out on two public data sets, namely Dredze and SpamArchive. The results show that the classification accuracy improves as the level of MSR increases, outperforming the accuracy of level 0 that relies on global features alone. Meanwhile, the ensemble of MSR improved the accuracy of MSR and outperformed all four MSR models for both datasets.