A Statistical Prior for Photo Forensics: Object Removal

If we consider photo forensics within a Bayesian framework, then the probability that an image has been manipulated given the results of a forensic test can be expressed as a product of a likelihood term (the probability of a forensic test detecting manipulation given that an image was manipulated) and a prior term (the probability that an image was manipulated). Despite the success of many forensic techniques, the incorporation of a statistical prior has not been previously considered. We describe a framework for incorporating statistical priors into any forensic analysis and specifically address the problem of quantifying the probability that a portion of an image is the result of content-aware fill, cloning, or some other form of information removal. We posit that the incorporation of such a prior will improve the overall accuracy of a broad range of forensic techniques.