A part-level learning strategy for JPEG image recompression detection

Recompression is a prevalent form of multimedia content manipulation. Different approaches have been developed to detect this kind of alteration for digital images of well-known JPEG format. However, they are either limited in performance or complex. These problems may arise from different quality level options of JPEG compression standard and their combinations after recompression. Inspired from semantic and perceptual analyses, in this paper, we suggest a part-level middle-out learning strategy to detect double compression via an architecturally efficient classifier. We first demonstrate that singly and doubly compressed data with different JPEG coder settings lie in a feature space representation as a limited number of coherent clusters, called parts. To show this, we visualize behavior of a set of prominent Benford-based features. Then, by leveraging such discovered knowledge, we model the issue of double JPEG compression detection in the family of feature engineering-based approaches as a part-level classification problem to cover all possible JPEG quality level combinations. The proposed strategy exhibits low complexity and yet comparable performance in comparison to related methods in that family. For reproducibility, our codes are available upon request to fellows.