A novel spam image filtering framework with Multi-Label Classification

Gray images, which could reasonably be considered as either spam or ham by different recipients, present significant obstacles to conventional binary spam filtering systems. The inconsistent labels of gray images will inevitably deteriorate the overall filter performance. In this paper, we present a novel framework named BFMLC (Binary Filtering with Multi-Label Classification) to take both spam image filtering and user preferences into account. The BFMLC framework comprises two-stage classification tasks: the filter-oriented binary classification and user-oriented multi-label classification. A filter based on the BFMLC framework can not only discriminate spam images from ham images, but also classify spam image as several predefined topics. According to user preference settings on the client side, the specific spam images (gray images) are delivered to individuals. Moreover, the BFMLC framework can be generalized to deal with text, image or mixed emails. We implement a spam image filtering system based on the BFMLC framework and conduct experiments in public personal datasets. The experimental results show that the system can identify spam images with the average accuracy of 96.309% and classify spam images as predefined topics with the average precision of 89.42%.