Digital image steganalysis using entropy driven deep neural network

Context-aware steganography techniques are quite popular due to their robustness. However, steganography techniques are misused to hide inappropriate information in some occurrences. In this paper, a new entropy-driven convolutional neural network is proposed to detect a stego-image. The proposed steganalysis method divides images into multiple sub-regions, and the highest entropy sub-regions are utilized for training the network. Small block size is used to eliminate the requirement of a pooling layer and to intact the flow of information content between the layers. A pooling layer is commonly used between the layers to reduce the size of the block at the cost of some information loss. The proposed method uses only sixteen non-trainable 3 × 3 size kernels, rather than thirty 3 × 3 and 5 × 5 size kernels used in the previous networks. In the proposed method, one global average pooling layer is considered to boost the performance at the last part of the network. The proposed method reduces the computational cost and improves detection accuracy. The proposed scheme is verified in the experimental analysis on the content-aware steganography algorithms, i.e., WOW, S-UNIWARD, and HILL, with multiple payloads. Two publicly available databases, i.e., BOWS2 and BOSSBase, are used to create numerous test scenarios.