Experimental Approach Based on Ensemble and Frequent Itemsets Mining for Image Spam Filtering

Abstract

Excessive amounts of image spam cause many problems to e-mail users. Since image spam is difficult to detect using conventional text-based spam approach, various image processing techniques have been proposed. In this paper, we present an ensemble method using frequent itemset mining (FIM) for filtering image spam. Despite the fact that FIM techniques are well established in data mining, it is not commonly used in the ensemble method. In order to obtain a good filtering performance, a SIFT descriptor is used since it is widely known as effective image descriptors. K-mean clustering is applied to the SIFT keypoints which produce a visual codebook. The bag-of-word (BOW) feature vectors for each image is generated using a hard bag-of-features (HBOF) approach. FIM descriptors are obtained from the frequent itemsets of the BOW feature vectors. We combine BOW, FIM with another three different feature selections, namely Information Gain (IG), Symmetrical Uncertainty (SU) and Chi Square (CS) with a Spatial Pyramid in an ensemble method. We have performed experiments on Dredze and SpamArchive datasets. The results show that our ensemble that uses the frequent itemsets mining has significantly outperform the traditional BOW and naive approach that combines all descriptors directly in a very large single input vector.