JPEG compression is a common standard for storing images. Its intrinsic lossy compression scheme, helps forensic analysts to disclose the history of the image by analyzing the histogram of discrete cosine transform (DCT) coefficients. In this paper, we propose a novel method based on sparse representation of the image, over learned over-complete dictionaries to conceal the statistical evidences of forgeries in JPEG compressed images. We will show that the proposed algorithm successfully fools currently used forensic methods. So that some of the state-of-the-art forensic tools which have an average performance accuracy of more than %90, produce random guess results on our anti-forensically doctored images.