An Incremental Learning Based Framework for Image Spam Filtering

Nowadays, an image spam is an unsolved problem because of two reasons. One is due to the diversity of spamming tricks. The other reason is due to the evolving nature of image spam. As new spam constantly emerging, filters’ effectiveness drops over time. In this paper, we present an effective anti-spam approach to solve the two problems. First, a novel clustering filter is proposed. By exploring the density-based clustering algorithm, the proposed filter is robust to spamming tricks. Then, we present a hierarchical framework by combining the clustering filter with other machine learning based classifiers to further improve the filtering capacity. Moreover, incremental learning mechanism is integrated to ensure the proposed framework be capable of adjusting itself to overcome new image spamming tricks. We evaluate the proposed framework on two public spam corpora. The experiment results show that the proposed framework achieves high precision along with low false positive rate.