From Forgery to Authenticity: Image Anti-Forensics via Reconstruction and Artefact Elimination

(1) I don't understand what the author is trying to say at each step in Figure 1. For example, what is the green dot and why does it disappear? Does the decision boundary of detector A refer to the curve in the center or the green dot? The authors did not express their work well with Figure 1.

(2) Authors should not spend nearly 2 pages in a conference article describing related work. Authors should focus on the presentation of the proposed methodology.

(3) I also can't understand the meaning expressed by Stage 2 in Fig. 2. Does the low pass filter in the bottom refer to the diffusion process at the top? If so, why is the diffusion process called a low pass filter? If not, what is the relationship between the top box and the bottom box?

(4) As shown in Figure 2, the author's method relies heavily on the effect of the initial detector. Does this mean that if the detector cannot locate the location of the tampered area, the author's method fails?

• Some implementation details and protocols provided in the paper is unclear

  • The term 'localization' is mentioned multiple times in the paper, but in mainstream research contexts, such as TruFor [(Guillaro et al., 2023)], 'detection' is generally considered as image classification (image-level), and 'localization' as image segmentation (pixel-level). However, the metrics in Table 1 are extremely high, so I'm not entirely sure whether these are the results of detection or localization; the paper does not make this clear.
  • Additionally, the description of the specific datasets and how the experiments were conducted is unclear. For instance, there is no information on which dataset was used for training, which for testing, or whether real images were included, etc. (lines 357–360).

• The structure of Stage 2 is relatively simple and is derived from two previous works, leading to a lack of overall technical contributions in this paper.

• Why aren’t TruFor and MVSS-net included as comparisons in Figures 4 and 5? Further, I kindly suggest changing these tables to separate a figure for each AUC, F1, and IOU, each figure contains five curves representing each model. The current figure putting AUC, F1, and IOU on the same chart is unreasonable.

Although the idea is interesting, the contribution of this work is generally incremental. The framework of GUIDE is similar to RePaint [1]. By comparing the Algorithms part between GUIDE and RePaint, the primary modification seems to be adding a frequency module. And TRM only comprises a prior FBCNN and the common Blur & Sharp operation.

To convincingly demonstrate the effectiveness of this contribution, it would be better to add a comparative analysis with the original RePaint and the proposed GUIDE, such as implementing RePaint + TRM settings for comparison.\

[1] RePaint: Inpainting using denoising diffusion probabilistic models. CVPR 2022

1. The motivation for developing an anti-forensics framework is not adequately discussed in the introduction. The authors should provide a thorough analysis and clear justification of the research purpose and motivations.

2. At the end of the introduction, the authors mention analyzing the trade-off between anti-forensics performance and image quality. However, intuitively, higher image quality should enhance anti-forensics effectiveness from a human perspective. This should be better substantiated in the manuscript.

3. The authors elaborate on the detailed diffusion process in Section 2.3, but this is not particularly informative and should not be included in the related works section.

4. The novelty of the proposed framework appears limited. Although the authors combine a diffusion-based module with a traditional texture refinement module to create anti-forensics images, the motivation behind this combination remains unclear.

5. As shown in Table 1, the proposed method's experimental results are suboptimal across different datasets and detectors. Additional datasets, such as Columbia, CASIAv1, and Carvalho, should be included to make the experiments more comprehensive.

6. This work does not include state-of-the-art (SOTA) detectors. The latest detection model used in experiments is Trufor, introduced at CVPR 2023.

In Figure 1, page 4 and Section 2.3, there are several equations which start with x and other symbols. However, many of these symbols do not seem to be defined anywhere in the paper. Some of them are defined but many are not. These are basic requirements in a technical paper and need to be addressed.

From the paper, it appears that the proposed approach build upon Diff-cfg (as mentioned inn the paper). It is not fully clear how much better the proposed approach is in comparison? It will be helpful to have a section which demonstrates what exactly is different in this paper and how it helps the overall performance. Though some of these details are discussed, it is not fully obvious from the paper.

Table 1 and 2 are crowded. It is difficult to see and infer how the proposed approach is actually performing in comparison to others. It is better to split these tables and show only the best cases of other detectors and show how the proposed approach compares to those. The detailed results could be documented in the supplementary section.