Arbitrary-Sized JPEG Steganalysis Based on Fully Convolutional Network

Arbitrary-Sized JPEG Steganalysis Based on Fully Convolutional Network

연구 분야: Strategies

학회: International Workshop on Digital Watermarking

Steganography detectors based on convolutional neural networks have achieved significant performance in recently years. Most existing networks, however, only focus on the detection performance of fixed-size images, and the sizes of training and testing images are usually 256 256 and 512 512. It can not meet the practical requirement of detecting arbitrary-sized steganographic images. Furthermore, there are few efficient methods for arbitrary-sized JPEG steganalysis. In this paper, we proposed a novel end-to-end steganalyzer based on fully convolutional network to address this issue. The characteristic of only containing convolutional layers allows the network to train and test arbitrary-sized images. In addition, the U-shaped network design can make element-wise classification, which provides state-of-the-art detection accuracy for both fixed and arbitrary size. The experimental results on standard image sources BOSSBase 1.01 and ALASKA #2 show that the proposed network can achieves superior performance.