On the impact of AI-based compression on deep learning-based source social network identification

Recognition of the social network of origin of an image is a relatively recent topic that is part of the techniques that fall under the umbrella of digital image forensics. It consists of the classification of images according to the social network on which they were posted. In contrast with other topics of digital image forensics, there are no works addressing counter forensic for source social network identification. Thus, we analyse the impact of image manipulations on its performances. We focus our study on AI-based compression, which tends to become the new compression solution with the upcoming standard JPEG AI. To conduct a fair analysis, we compare the AI-based compression with the conventional legacy JPEG compression, and also include three other manipulations: median filtering, Gaussian blurring, and additional white Gaussian noise, which are often used to assess the robustness of digital image forensic methods. We define two sets of parameters based on the resulting image quality in terms of structural similarity, which correspond respectively

to attacks with strong and limited image degradation. In the context of strong downgrade of the image quality, all the manipulations lead to similar decrease in performance, while

for attacks that preserve ima ge quality, AI-based compression is able to reach a drop in identification rate twice higher than the other manipulations.