A Sanity Check for AI-generated Image Detection

We appreciated the insightful reviews and valuable suggestion. We hope the following response can address your concerns.

Q1. The combination of patch-level features and global features proposed is not novel.

Although low-level information and global features are common, we have carefully designed components to address significant challenges in AI-generated image detection, which have been verified by many benchmarks, e.g., AIGCDetectBenchmark, GenImage, and Chameleon. Our research contributions are fourfold:

1. Patchwise Feature Extraction. We designed a block scoring mechanism to selectively highlight the highest and lowest frequency blocks. Our patchwise feature extraction module targets disparities in underlying information, such as the presence of white noise during image capturing.

2. Semantic Feature Embedding. Our semantic feature embedding module captures object co-occurrence and contextual relationships. For instance, penguins are unlikely to appear on the grasslands of Africa.

3. Generalization on different settings. The significant improvement over previous SOTA methods in various settings demonstrates our method's strong generalization capability.

4. Visualization. As visualized in Fig 4 of main paper, without semantic feature extraction, which results in many images with evident semantic errors going undetected. Similarly, without patchwise feature extraction, numerous images lacking semantic errors still contain differing underlying information that remains unrecognized.

Therefore, we believe that our method is novel and valuable. In addition, we present a more challenging dataset characterized by comprehensive manual annotation, real-world applicability, high resolution, and large scale.

Q2. The comparsion of recent works.

Thanks for the suggestion. We have included new baseline model, namely, FatFormer, and use their publicly available model for evaluation. The detailed comparison is as follows and we will add them in revised version.

Notably: LaRE². However, we discovered that the LaRE² method was open-sourced only in the last three weeks and lacks training weights. In our attempts to replicate the results, we encountered missing essential code files, an issue that has also been noted by others in the GitHub issues section.

The table below shows the results of FatFormer and AIDE on AIGCDetectBenchmark. As shown in the table below, AIDE surpasses FatFormer by a large margin.

Q3. The inference throughput of the proposed method is not provided, nor is it compared with previous methods.

Thanks for pointing this out. We further evaluate throughput of the proposed method and recent works.

Q4. Whether the performance increase claimed in the article is due to increased FLOPS/PARAMS or the AIDE itself.

We further conducted additional experiments to explore the core components of AIDE.

1. Exploring the Patchwise Feature Extraction in our AIDE: We replaced the high and low-frequency blocks with the original images and conducted experiments on AIGCDetectBenchmark. As shown below, our high and low-frequency blocks are extremely effective.

2. Ablation Study. The CLIP-ConvNeXt occupies 94% of the parameters in AIDE. When using only CLIP-ConvNeXt, there is a significant drop in performance on AIGCDetectBenchmark, as shown below. This demonstrates that our method's effectiveness is not due to an increase in the number of parameters.

We sincerely appreciate your careful reviews and constructive comments. We hope our response can address your concerns.

Q1. Theoretical explanation for the proposed method.

1. High Frequency Patches. Generative models often produce smoother patches in high-frequency areas, such as when generating hair. As a result, image patches in these regions can serve as a powerful tool for identifying AI-generated images.

2. Low Frequency Patches. In low-frequency areas, like patches of solid color, a generative model primarily needs to match the surrounding color. However, when a camera captures real images, achieving perfectly uniform color is very unlikely due to the presence of various types of noise.

Q2. Comparing the entire dataset with subsets of the AIGCDetectBenchmark dataset seems unreasonable.

Sorry for the confusion. We do not mean to compare our dataset with existing ones. Our goal is to show that our dataset serves as a valuable extension of existing AI-generated image detection evaluation.

1. Deceptively real. As far as we know, this is the first dataset containing images that are manually curated, which are indistinguishable from human observers.
2. Diversity. The Chameleon dataset is derived from real-world scenarios, ensuring a broad and authentic representation.
3. High resolution. Considering that the majority of images have resolutions surpassing 720P and reaching up to *4K, the entire dataset exhibits exceptional clarity. And this high resolution poses increased challenges for the identification of artifacts.
4. Large-Scale. As illustrated in Tab 2 of the main paper, our current Chameleon dataset stands out as the largest available.

Consequently, both high- and low-frequency patches are valuable in the detection of AI-generated images.

Q3. It may be beneficial to employ Grad-CAM to visualize the regions of interest when detecting forgery in images.

Thanks for the advice. We have incorporated new visualization in Figure 6, located in the appendix of the revised manuscript. Additionally, we have included a more comprehensive analysis in L941–949 of the main paper.

We sincerely appreciate your insightful reviews and positive comments. We hope our response can address your concerns.

Q1. Given the limited number of samples and categories, how can the authors ensure whether the images are representative enough.

Thank you for pointing that out. We will include more analysis in the revised version.

1. Our dataset actually consists of diverse categories from four parent categories, namely, animal, human, object, and scene. For example, the animal category includes more specific subcategories, such as cats, tigers, dogs, and others.

2. Our dataset was constructed with the goal of simulating real-world scenarios, to focus the images indistinguishable from humans. We agree that the animal category contains fewer images that the other ones, this actually reflects the flaw of modern generative models or human perception systems, i.e., the generated animal images are more easily detected, while the other categories are relatively more difficult.

3. As illustrated in Fig 1 of the main paper, our dataset features a diverse array of categories. Unlike previous test sets that focus on relatively singular scenarios, our dataset emphasizes real-world scenarios.

4. Although our collection of fake animal images comprises only a few hundred photos, they encompass a broad range, as visualized in Fig 1 of the main paper, and effectively pass the “Turing test.” Additionally, we assessed some recent methods through a fine-grained experiment, which evaluates accuracy across the four main categories, with results shown in the accompanying tables. We present the “fake image Acc / real image Acc” for detailed analysis. As demonstrated, the performance of the animal category is comparable to that of other categories. This indicates that our data is sufficient for validating detection methods.

Therefore, the images are representative enough for validating AI-generated image detectors.

Q2. CLIPMoLE[Liu et al., 2024] is supposed to discuss the major differences to AIDE, and serve as a possibly stronger baseline.

Q2.1 The major differences to AIDE.

Thanks for pointing this out. Our AIDE differs from CLIPMoLE in three aspects:

1. Motivation. CLIPMoLE's mixture-of-low-ranks is specifically designed for parameter-efficient fine-tuning. As mentioned in L71–75 of the main paper, AIDE is primarily developed for identifying AI-generated images originating from multiple factors. These cues may intuitively arise from various sources, including low-level textures, pixel statistics, and semantic cues.

2. Methodology. CLIPMoLE adapts only the MLP layers of deeper ViT blocks through a combination of shared and separate LoRAs within an MoE-based structure. As outlined in L76–80 of the main paper, AIDE incorporates a scoring module to capture low-level pixel statistics and harness global semantics with CLIP. The fused features are then utilized for AI-generated image detection.

3. Features. CLIPMoLE is limited to encoding image features. In contrast, AIDE extracts both the underlying noise features and high-level semantic features.

Q2.2 CLIPMoLE[Liu et al., 2024] serves as a possibly stronger baseline.

Thanks for the suggestion. The codes and models have not been released yet, we will incorporate them as a baseline if the codes are available, as promised in their paper.

Q3. Evaluate JPEG compression with more realistic quality factors e.g. QF=75, QF=50 and Gaussian blur for sigma=3 and 4 on AIDE.

Thanks for pointing this out. We have further conducted the robustness of AIDE on AIGCDetectBenchmark as you suggested.

The results from the two robustness experiment tables clearly demonstrate that our method exhibits a high degree of robustness.

Q4. Regarding the dataset construction, have the authors strictly obey the 4 principles as presented in p4.

Yes, we strictly adhered to our proposed guidelines when constructing the dataset. We did not use any model to validate the dataset's performance beforehand. Instead, after manual annotation, we directly included the images that humans failed to distinguish their sources, i.e., from camera or generative model.

I thank the authors for their rebuttal to help alleviate some of my concerns. A main contribution of this work is to create the Chameleon dataset, however, I didn't find the dataset or a link to it after having checked the paper and its supplementary material.

As a paper with a proposed dataset, I believe it is very important and SUPPOSED to release the dataset to the public (should be anonymously at the review phase). Otherwise, the community can't reproduce the results and conduct their experiments on this dataset. I'm looking forward to this proposed dataset and will adjust my rating score depending on its availability.

Besides, I strongly suggest the authors reflect rebuttal responses in the revised paper, e.g., including discussions on the MoE detector [1] in related works, adding results on QF=75, QF=50 and sigma=3 and 4 in the main body.

Thank the authors for releasing the dataset and for their efforts in revising the paper. Regarding the dataset, a minor suggestion is to organize it in a more structured way, e.g. adding subfolders containing the four categories in the fake folder (currently, all fake images were put in the same folder), which I believe will help the community do more fine-grained analysis conveniently.

Besides, while I tend to accept this work, I feel a bit concerned regarding the high imbalance of fake images gathered from some categories, e.g. only 313 fake images from the Animal category. I hope the authors could add a discussion in the final version regarding the scalability of the four rules, i.e. discussing the feasibility of increasing the number of certain-category fake images, or explaining in the revised paper the main reason accounting for such high imbalance between categories. Otherwise, it seems quite weird at the moment.

We sincerely appreciate constructive suggestions. We hope the following response can address your concerns.

Q1. Undertaking an in-depth analysis of Chameleon dataset.

Thanks for the suggestion. We have included a detailed discussion and analysis in L58-63 and L256-261 of the main paper. The unique contribution of our Chameleon lies in the following four aspects:

1. Deceptively real. All AI-generated images within the dataset are manually labeled and have successfully deceived human perception.
2. Diversity. The Chameleon dataset is derived from real-world scenarios, ensuring a broad and authentic representation.
3. High resolution. Considering that the majority of images have resolutions surpassing 720P and reaching up to 4K, the entire dataset exhibits exceptional clarity. And this high resolution poses increased challenges for the identification of artifacts.
4. Large-Scale. As illustrated in Tab 2 of the main paper, our current Chameleon dataset stands out as the largest available.

Q2. It would be good if the authors collect fake images of fake news on the Internet and conduct experiments.

Thanks for the suggestion. We did not source fake images from fake news on purpose, and we believe this can be treated as a very meaningful future work. However,

1. As discussed in L206~210 of the main paper, our Chameleon dataset is derived from a wide range of sources, some of them may indeed be from fake news.

2. We believe the techniques used for creating fake images in the fake news are also: Photoshop and diffusion models, which has been covered by images in our datasets, thus, the models developed for detecting these fake images should be equally applicable to ones in fake news.

We will clarify these aspects in the revised version.

We highly appreciate your recognition of our work and are encouraged by your positive comments. We hope our response can address your concerns.

Q1. Some validation of the new dataset in its visual quality perspective will make the work stronger. E.g., by human studies and comparisons between datasets.

Thank you for the advice. When developing the Chameleon dataset, we devoted significant effort to addressing the issues found in previous datasets, as depicted in Fig. 1 of the main paper. Additionally, we invested substantial manpower in manually annotating existing datasets.

1. Visualization. Fig. 1 in the main paper illustrates a comparison between our dataset and other existing datasets, clearly demonstrating that our dataset offers a more realistic portrayal.

2. Manual Annotation. Chameleon is the first dataset with manual annotations. As written in L247–249 of the main paper, our dataset is manually labeled by having each image evaluated by two annotators. We retain the images only when both annotators mis-classify them as real. Notably, all AI-generated images included in the dataset have successfully tricked human perception.

3. As per your suggestion, we intend to conduct a randomized selection of 500 images from the existing dataset for evaluation by annotators. This assessment will focus on various aspects, such as the visual quality of images, the depiction of scenes, and the determination of whether the images are authentic or generated by artificial intelligence.

Q2. In Table 6, it is better to also list the accuracy numbers under no perturbations for convenient comparison.

Thanks for the suggestion. We have incorporated it in the revised manuscript.

Q3. More recent related work can be compared or discussed.

Thank you for highlighting this. We selected another recent approach for comparison, namely, FatFormer [1]. We have incorporated the discussion in the related work in our paper.

We take the publicly available FatFormer model. The table below shows the results of FatFormer and AIDE on AIGCDetectBenchmark. As shown in the table below, AIDE surpasses FatFormer by a large margin.

[1] Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection. In CVPR, 2024. (arXiv preprint arXiv: 2312.16649)