Abstract
This paper proposes a method for still image steganalysis using noise estimation on sparse code. Sparse coding is one of linear representation methods, and has the property that only a small number of the components of basic images are significantly non-zero. Basic images are not fixed but adaptive for input data, unlike DCT decomposition. It is proven that the estimation of independent component analysis model for sparse data is equivalent to sparse coding, which decompose input an image data into basic images and its coefficients. In general, the coefficients of basic images are distributed non-Gaussian. We assume that the change of image by embedding as Gausian noise. Because sparse code shrinkage can effectively separate Gausian distribution from non-Gaussian distribution, we apply the estimated noise by sparse code shrinkage to image steganalysis. In the experiments, we show our method outperforms previous steganalysis methods: Farid, Goljan and Wang's method.
