Dual JPEG Compatibility: a Reliable and Explainable Tool for Image Forensics

TL;DR

This work tackles JPEG forgery detection by exploiting block-level compatibility under JPEG pipelines, introducing a dual extension of antecedent search that can handle arbitrary composed pipelines. It formalizes the forward and backward JPEG processes, derives theoretical bounds for constrained local search, and shows that modified blocks produce three mismatch types (grid, quantization, pipeline) that can be detected to localize forgeries at the JPEG-block level. In experiments, the approach outperforms three state-of-the-art deep learning baselines in idealized settings and achieves high block-level localization accuracy, especially for unmodified images or when second compressions strengthen the original quality factor. As a proof-of-concept, the method emphasizes explainability and deterministic detection, while noting limitations related to exact pipeline knowledge and computational cost, and outlining directions for robustness to unknown pipelines and faster search.

Abstract

Given a JPEG pipeline (compression or decompression), this paper demonstrates how to find the antecedent of an 8x8 block. If it exists, the block is considered compatible with the pipeline. For unaltered images, all blocks remain compatible with the original pipeline; however, for manipulated images, this is not necessarily true. This article provides a first demonstration of the potential of compatibility-based approaches for JPEG image forensics. It introduces a method to address the key challenge of finding a block antecedent in a high-dimensional space, relying on a local search algorithm with restrictions on the search space. We show that inpainting, copy-move, and splicing, when applied after JPEG compression, result in three distinct mismatch problems that can be detected. In particular, if the image is re-compressed after modification, the manipulation can be detected when the quality factor of the second compression is higher than that of the first. Through extensive experiments, we highlight the potential of this compatibility attack under varying degrees of assumptions. While our approach shows promising results-outperforming three state-of-the-art deep learning models in an idealized setting-it remains a proof of concept rather than an off-the-shelf forensic tool. Notably, with a perfect knowledge of the JPEG pipeline, our method guarantees zero false alarms in block-by-block localization, given sufficient computational power.

Figures (12)

• Figure 1: Demonstration of finding incompatible JPEG blocks. The chimney was removed using in-painting ( i.e. uncompressed data), and birds were spliced with different Quality Factors ($\textit{QF}_{\mathrm{}}$s), one of them is fully aligned with the JPEG grid but not the other. The window was copy-moved aligned on the JPEG grid but not fully aligned, i.e. portions of blocks belong to the original image.
• Figure 2: General flowchart of our method. For each block of the image obtained with a pipeline $\mathcal{F}$, we search an antecedent with a local search constrained with upper bound $M_{\mathcal{F}}$. If the block is incompatible, we can conclude that it was altered. If the block is unsolved, the decision is left to the agent depending on the number of iterations $N$.
• Figure 3: Probability for compatible blocks to be unsolved after a given number of iterations (FPR). This empirical curves were obtained by averaging the results of all single and double compression experiments ($\sim1$M blocks) from section \ref{['sec:comparison']}.
• Figure 4: General scenario for modifying a JPEG image in the spatial domain. Blues boxes represent blocks in the spatial domain (pixels) and orange boxes represent blocks in the frequency domain (DCT coefficients). $\mathcal{C}$ and $\mathcal{D}$ are the compressors and decompressors used. $\mathbf{Q}$ are the quantization tables. Modifications applied to $\mathbf{C}$ can be fully aligned (FA) or not fully aligned (NFA).
• Figure 5: Examples of fully aligned and not fully aligned modifications. Black lines represent blocks. Example \ref{['fig:NFA_trans']} illustrates block translations from one image to another, on the contrary example \ref{['fig:aligned_c']} illustrates a finer segmentation.