A fast image spam filter based on ORB

Image spam has become a new obfuscating method to bypass conventional text based spam filters. In this paper, a new kind of image spam filtering method is proposed based on the characteristics of the spam being sent repeatedly and their contents being highly resemble to each other. We extract sub-block color histogram and ORB as image feature and run a scalable vocabulary tree to detect image spam. The system is tested on Mark Dredze´s dataset and our own Chinese image spam corpus. Experimental results demonstrate that the proposed method can achieve good accuracy while having a less than 0.02% false positive rate.