Abstract
With the tremendous growth of the internet, cyberspace is facing several
threats from the attackers. Threats like spam emails account for
55\% of total emails according to the Symantec monthly
threat report. Over time, the attackers moved on to image spam to evade
the text-based spam filters. To deal with this, the researchers have
several machine learning and deep learning approaches that use various
features like metadata, color, shape, texture features. But the Deep
Convolutional Neural Network (DCNN) and transfer learning-based
pre-trained CNN models are not explored much for Image spam
classification. Therefore, in this work, 2 DCNN models along with few
pre-trained ImageNet architectures like VGG19, Xception are trained on 3
different datasets. The effect of employing a Cost-sensitive learning
approach to handle data imbalance is also studied. Some of the proposed
models in this work achieves an accuracy up to 99\% with
zero false positive rate in best case.