A Statistical Modeling Framework for DCT Coefficients of Tampered JPEG Images and Forgery Localization

Various manipulations on JPEG images introduce single and multiple compression artifacts for forged and unmodified areas respectively. Based on the statistical analysis of JPEG compression cycle and on the finite mixture paradigm, we propose in this paper a modeling framework for AC DCT coefficients of such tampered JPEG images. Its accuracy is numerically assessed using the Kullback-Leibler divergence on the basis of a tampered JPEG image dataset built from six well-known uncompressed color image databases. To illustrate the framework utility, an application in image forgery localization is proposed. By formulating the localization as a clustering problem, we use the plug-in Bayes rule combined with a simple EM algorithm to distinguish between forged and unmodified areas. Numerous experiments show that, when the quality factor of final JPEG compression is high enough, the proposed modeling framework yields higher localization performances in terms of
$F_{1}$
-score than prior art regardless of divers local manipulations.