Diffusion models meet image counter-forensics

TL;DR

This work examines diffusion models as counter_forensics tools that erase forgery traces embedded in camera processing, thereby deceiving forensics detectors. It introduces Diffusion Counter-Forensics (Diff-CF) and its guided variant Diff-CFG, which diffuse forged images up to a time_step and then reverse_diffuse to purify, with optional guidance to stay close to the input. The study shows diffusion_based purification often outperforms existing counter_forensic methods in removing forensic traces while preserving perceptual quality, across multiple detectors and datasets, albeit with tradeoffs controlled by time_step and guidance. The findings highlight both the potential for authenticating image integrity and the need for developing detection methods robust to diffusion_purification, motivating future work on trace_analysis and counter_forensic defenses.

Abstract

From its acquisition in the camera sensors to its storage, different operations are performed to generate the final image. This pipeline imprints specific traces into the image to form a natural watermark. Tampering with an image disturbs these traces; these disruptions are clues that are used by most methods to detect and locate forgeries. In this article, we assess the capabilities of diffusion models to erase the traces left by forgers and, therefore, deceive forensics methods. Such an approach has been recently introduced for adversarial purification, achieving significant performance. We show that diffusion purification methods are well suited for counter-forensics tasks. Such approaches outperform already existing counter-forensics techniques both in deceiving forensics methods and in preserving the natural look of the purified images. The source code is publicly available at https://github.com/mtailanian/diff-cf.

Figures (4)

• Figure 1: Illustration of using diffusion models as a counter-forensic technique. A forged image from FAU dataset FAU_dataset, correctly detected by ZERO ZERO, produces no detection after diffusion purification.
• Figure 2: Results obtained by different forensics methods on the different versions of image r7710a7fat from the Korus dataset korus1korus2. We observe that Choi choi, ManTraNet mantranet, and Noiseprint noiseprint feature no detection when Diff-CF or Diff-CFG are applied. For Splicebuster splicebuster2015 and TruFor TruFor, even if counter-forensics techniques are not completely able to deceive them, the proposed approaches degrade their detections the most. More examples are included in the supplementary materials.
• Figure 3: Image quality comparison for all considered counter-forensics methods. We observe that both Diff-CF and Diff-CFG are good at preserving the fine textures and edges of the image while CamTE and BM3D blur all these fine structures.
• Figure 4: Study of the impact of the time-step $t^*$ (left-hand side), and guidance scale $s$ (right-hand side). For each parameter, we evaluate its influence on the forgery traces removal task (top) and on the purified image quality (bottom). For the forgery traces removal task, we plot the average difference between the performance before and after purification for the best-performing methods in the original dataset as a function of the parameters' value. The colored background area represents the 95% confidence interval. For the Image quality assessment, all five metrics presented in Sec. \ref{['subsec:qualityassesment']} are plotted as a function of the parameters' value, each with a different axis, for better visualization. This figure is best viewed in color. An interactive version of these plots is included in the supplementary material.