Based on a local stationary model of images, we propose a locally adaptive method to combat the least significant bit(LSB) matching steganography. We model the LSB matching embedding as additive Gaussian noise, use the Gaussian mixture model to describe local detail components, and estimate the model parameters with the expectation maximization algorithm. The smallest variance is selected as an estimation of
local stego-noise variance. The weighted sum features of the local variance histogram are extracted to characterize changes in regions with different complexity between the cover and stego images. Features extracted from an image and its down-sampled version are combined and sent into a classifier. The experimental results on two sets of uncompressed images show that the proposed steganalyzer outperforms the prior art and provides
reliable results for embedding rates as low as 0.25 bits per pixel.