Compression Robustness of Digital Image Forensics Evaluating Demosaicing Consistency

Image tampering detection techniques based on the analysis of CFA-demosaicing traces excel by their detection accuracy, however, they are very sensitive against slightest changes in an image. Therefore, their robustness against lossy compression schemes is a crucial issue with respect to their applicability. The robustness of seven different tampering detection techniques of this type is experimentally assessed considering eight distinct lossy compression schemes, including four ISO/IEC JPEG standards as well as four algorithms originating from video intra-frame coding schemes. Overall, we find that the accuracy of the forensic analysis is reduced for increasing compression strength as expected, however, the relative performance of the compression schemes is different for the two considered test data sets and the tampering detection algorithms, respectively. In any case, distinct compression algorithms change the ranking of the forensic algorithms in terms of tampering detection accuracy in different manner across varying compression strengths and thus, it is almost impossible to provide general guidelines for compression algorithm deployment. With respect to compression robustness, a recent neural network-based technique predicting the relative positions of CFA blocks in the image turns out to be most robust against any type of compression, but is already best performing on uncompressed data.