Steganalysis using Noise Variance Estimation

Steganography is the art and science of hiding information in an innocuous medium. Digital imagery is a medium with specific noise characteristics. The devices used to capture a digital image, such as the charge coupled device (CCD) in a digital camera is designed to have a relatively small noise characteristic. This noise characteristic is often reduced by the compression used such as the JPEG standard, resulting in a very low noise media. The addition of steganography to the image has the effect of introducing changes to this "natural" image noise. This paper presents an approach to estimate and model the noise present in an image. Using this estimation, it is shown how steganography introduces detectable changes to this natural noise. This approach is demonstrated on three freely available but difficult to detect embedding techniques, F5, JSteg, and Model-based embedding, and show that it results in features that serve as statistically significant discriminators.