Understanding multi-layer perceptrons on spatial image steganalysis features
Lilei Zheng, Ying Zhang, Vrizlynn L. L. Thing · 2017
This paper proposes to apply single layer perceptrons (SLP) and multi-layer perceptrons (MLP) for image steganalysis. Compared to the popular ensemble classifier (EC) that previous works used, the MLP has larger capacity to fit the training data and improves the test performance accordingly. When we compare different classifiers, the overfitting gap between the training and test results is a good indicator of a classifier's learning ability. On the BOSSbase V1.01 dataset, the experiments demonstrate that the MLP is able to learn a better feature representation than the raw input feature. For instance, on one of the spatial image steganalysis features, a relative performance gain of 2.62% (from 72.58% to 74.48%) is obtained by the MLP over the EC.