Size-Independent Reliable CNN for RJCA Steganalysis

Detection of image steganography is principally implemented with supervised machine learning detectors. There are two main drawbacks to this approach: the detectors are overly specific to a given image source, and the performance guarantees are only empirical. In this work, we further study a previously proposed deep learning detector that exploits natural image structure imposed by JPEG compression with high quality. We show in a controlled environment that for a fixed JPEG compressor, the soft outputs of a deep learning classifier - the logits - follow a Gaussian distribution. We prove a scaling law stating that the variance of this distribution scales linearly with the image size. By disabling padding in the convolutional neural network, we demonstrate that the mean of the logit distribution does not change, allowing us to directly analyze images of different sizes. Focusing on the logits, we show that we can prescribe a threshold with a theoretical false positive rate for a wide range of image sizes, which is then closely satisfied on real cover images, even for small probabilities such as 10−4. Moreover, the detection power on steganographic images still generalizes to non-adaptive and content-adaptive steganography.