Three Forensic Cues for JPEG AI Images

TL;DR

JPEG AI introduces new forensic traces that render traditional JPEG forensics less effective and risk confusing AI compression with synthetic-generation artifacts. The authors offer three analytic cues—color-channel correlations, rate-distortion-based recompression features, and latent-space quantization artifacts—to detect JPEG AI compression, identify recompression, and distinguish JPEG AI from AI-generated images. Empirical results on real and synthetic data show strong performance and generalization across encoders and bitrates, with interpretable features suitable for forensic analysts. This work provides a practical, interpretable foundation for JPEG AI forensics and motivates further research into compression-forensics for AI-based codecs.

Abstract

The JPEG standard was vastly successful. Currently, the first AI-based compression method ``JPEG AI'' will be standardized. JPEG AI brings remarkable benefits. JPEG AI images exhibit impressive image quality at bitrates that are an order of magnitude lower than images compressed with traditional JPEG. However, forensic analysis of JPEG AI has to be completely re-thought: forensic tools for traditional JPEG do not transfer to JPEG AI, and artifacts from JPEG AI are easily confused with artifacts from artificially generated images (``DeepFakes''). This creates a need for novel forensic approaches to detection and distinction of JPEG AI images. In this work, we make a first step towards a forensic JPEG AI toolset. We propose three cues for forensic algorithms for JPEG AI. These algorithms address three forensic questions: first, we show that the JPEG AI preprocessing introduces correlations in the color channels that do not occur in uncompressed images. Second, we show that repeated compression of JPEG AI images leads to diminishing distortion differences. This can be used to detect recompression, in a spirit similar to some classic JPEG forensics methods. Third, we show that the quantization of JPEG AI images in the latent space can be used to distinguish real images with JPEG AI compression from synthetically generated images. The proposed methods are interpretable for a forensic analyst, and we hope that they inspire further research in the forensics of AI-compressed images.

Figures (8)

• Figure 1: Color correlations for the three color channels. Uncompressed images (violet) exhibit lower correlations than JPEG AI. The correlations are higher for stronger compression (corresponding to lower bitrates).
• Figure 2: Distinctiveness of color correlations, shown on the example of $\rho(\mathbf{r})$: (a) reference distributions uncompressed and JPEG AI, as on the right of Fig. \ref{['fig:color_space_feature']}. (b) JPEG AI color conversion and 4:2:0 chroma subsampling on uncompressed images (pink) notably increases the correlations over uncompressed images (violet). (c) JPEG compression qualities 20 to 100 slightly increase correlations. (d) The synthetic image generators Midjourney-V5 and Firefly slightly increase correlations.
• Figure 3: PSNR curve between original image and compression $k$ per bitrate. We analyze JPEG AI (left), the AI codec by Ballé et al.Balle_Var_Autoencoder (middle) and JPEG (right). See text for details.
• Figure 4: PSNR between compression $k-1$ and $k$ per bitrate. We analyze JPEG AI (left), the AI codec by Ballé et al.Balle_Var_Autoencoder (middle) and JPEG (right). See text for details.
• Figure 5: Dependencies between rate $r(k)$ and PSNR $p_{\mathrm{inc}}(k)$ during $k$ compression runs to differentiate between single and recompressed JPEG AI images. In particular, $r(1)$ and the difference $r(3)-r(2)$ are more discriminative than the others.