Resumen
Digital images are very popular and commonly used for hiding crucial data. In a few instances, image steganography is misused for communicating with improper data. In this paper, a robust deep neural network is proposed for the identification of content-adaptive image steganography schemes. Multiple novel strategies are applied to improve detection performance. Two non-trainable convolutional layers is used to guide the proposed CNN with fixed kernels. Thirty-one kernels are used in both non-trainable layers, of which thirty are high-pass kernels and one is the neutral kernel. The layer-specific learning rate is applied for each layer. ReLU with customized thresholding is applied to achieve better performance. In the proposed method, image down-sampling is not performed; only the global average pooling layer is considered in the last part of the network. The experimental results are verified on BOWS2 and BOSSBase image sets. Content-adaptive steganography schemes, such as HILL, Mi-POD, S-UNIWARD, and WOW, are considered for generating the stego images with different payloads. In experimental analysis, the proposed scheme is compared with some of the latest schemes, where the proposed scheme outperforms other state-of-the-art techniques in the most cases.