Deep learning hierarchical representations for image steganalysis

원문 초록

Nowadays, the prevailing detectors of steganographic communication in digital images mainly consist of three steps, i.e., residual computation, feature extraction, and binary classification. In this paper, we present an alternative approach to steganalysis of digital images based on convolutional neural network (CNN), which is shown to be able to well replicate and optimize these key steps in a unified framework and learn hierarchical representations directly from raw images. The proposed CNN has a quite different structure from the ones used in conventional computer vision tasks. Rather than a random strategy, the weights in the first layer of the proposed CNN are initialized with the basic high-pass filter set used in the calculation of residual maps in a spatial rich model (SRM), which acts as a regularizer to suppress the image content effectively. To better capture the structure of embedding signals, which usually have extremely low SNR (stego signal to image content), a new activation function called a truncated linear unit is adopted in our CNN model. Finally, we further boost the performance of the proposed CNN-based steganalyzer by incorporating the knowledge of selection channel. Three state-of-the-art steganographic algorithms in spatial domain, e.g., WOW, S-UNIWARD, and HILL, are used to evaluate the effectiveness of our model. Compared to SRM and its selection-channel-aware variant maxSRMd2, our model achieves superior performance across all tested algorithms for a wide variety of payloads.