acceptodds
Under review as a conference paper at ICLR 2027

Probing AI-Generated Image Detectors via Black-Box Anti-Forensics on Semantic and Artifact Cues

Abstract

Recent advances in generative models have made AI-generated images increasingly difficult to distinguish from real ones, motivating detectors that exploit semantic representations from vision foundation models (VFMs) and synthesis-specific artifacts left by generative models. However, it remains unclear whether these cues capture intrinsic evidence of forgery or merely reflect incidental characteristics of the generation process. In this work, we investigate this question through systematic anti-forensic probing, finding that detectors can be substantially compromised by independently manipulating these two types of evidence: the semantic representations used for discrimination can be steered toward authentic semantics, while synthesis-specific artifacts can be suppressed without noticeably altering image content. Based on these findings, we propose CRAS, a black-box anti-forensic framework that jointly targets semantic and artifact evidence. Specifically, a cross-modal reconstruction module steers forged images embeddings toward authentic text anchors in the VFM feature space, followed by an artifact suppression module that reduces the synthesis-specific artifacts left by the reconstruction process. Extensive experiments show that CRAS achieves the best anti-forensic performance among existing methods while preserving high visual fidelity. Our results suggest that current AIGI detectors rely heavily on both types of evidence that are not invariant to carefully controlled transformations, highlighting the importance of learning manipulation-invariant forensic evidence.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Anti-Forensics, AIGI Detection, Stable Diffusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yp7KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CRAS, a black-box anti-forensic framework for AI-generated image detectors. The method combines cross-modal reconstruction to manipulate semantic cues with artifact suppression to reduce synthesis-specific traces. Experiments across four datasets and ten detectors demonstrate strong attack effectiveness and high visual fidelity, supported by comparisons with existing methods and ablation studies.

Critical strengths
  1. Meaningful research problem. The paper investigates the robustness of AI-generated image detectors by examining their reliance on semantic and synthesis-specific artifact cues. This offers a useful perspective on understanding detector vulnerabilities.

  2. Comprehensive experimental evaluation. The method is evaluated across four datasets and ten detectors, covering both CLIP-based and conventional architectures. The results demonstrate strong attack effectiveness while maintaining good image quality, supporting the practical value of the proposed framework.

Critical weaknesses
  1. Insufficient validation of the underlying mechanism. The paper claims that CRM and ASM independently manipulate semantic and artifact cues, respectively. However, diffusion-based reconstruction may also alter low-level artifacts, while artifact suppression may affect semantic representations. The current ablation studies demonstrate the effectiveness of each module but do not sufficiently establish the claimed separation of these two mechanisms.

  2. Limited methodological novelty. Both diffusion-based latent optimization with CLIP guidance and encoder-decoder-based artifact suppression build on established techniques. The main contribution appears to be their combination, while the claimed complementary mechanisms are not rigorously validated. This weakens the novelty and contribution to understanding detector vulnerabilities.

Questions for authors
  1. Mechanism validation (Clarification). How can the authors establish that CRM primarily manipulates semantic cues while ASM specifically suppresses synthesis artifacts? Could the authors provide quantitative evidence showing how each module affects both types of features?

  2. Contribution of semantic optimization (Clarification). How much of CRM's attack effectiveness comes from the semantic alignment objective rather than the diffusion reconstruction process itself? A comparison with standard DDIM reconstruction without semantic optimization would help clarify the contribution of the proposed mechanism.

2 comments

We thank the reviewer for the feedback and have conducted additional experiments to address your questions.


Response to Weakness 1 & Q1

To quantify the effects of each module on semantic and artifact cues, we use the following methods:

  1. We extract features from the original images, CRM-processed images, and ASM-processed images using a pretrained CLIP model. For each module, we compute the mean cosine distance and mean L2 distance between the features of the processed images and those of the corresponding original images. Smaller distances indicate less change to semantic cues.
  2. To quantify changes in artifact cues, we separately compare CRM-processed and ASM-processed images with their corresponding original images, computing the Jensen–Shannon (JS) distance between power spectra and the RMSE between log-power spectra in the mid- and high-frequency bands.

Table 1: Effect of Each Module on Semantic Cues.

Mean Cosine DistanceMean L2 Distance (Normalized)
CRM0.4210.909
ASM0.0960.425

Analysis: The greater distance between the features of CRM-processed images and those of the original images indicates that CRM primarily modifies semantic cues.

Table 2: Effect of Each Module on Artifact Cues.

Mid-Frequency JS DistanceHigh-Frequency JS DistanceMid-Frequency RMSE (dB)High-Frequency RMSE (dB)
CRM0.4120.5958.13111.060
ASM0.3610.5257.5609.839

Analysis: The mid- and high-frequency power spectra of ASM-processed images are closer to those of the original images, indicating that ASM suppresses more artifacts.


Response to Weakness 2

We explain how CRM and ASM each differ from existing methods as follows:

  • Compared with existing CLIP-guided diffusion optimization methods, we designed two text pools containing the most representative real and fake prompts. These prompts achieve excellent performance in zero-shot detection with CLIP. We also designed a different alignment loss to make the CLIP embeddings of reconstructed images match the real prompts more strongly.
  • The technical differences between ASM and conventional encoder–decoder methods lie in the training pairs and how the model is used. We train ASM to map each DRCT Stable Diffusion reconstruction to its corresponding original MS-COCO image to suppress artifacts left by reconstruction. To adapt the model to the anti-forensic task, we modify its convolutional layers. During fine-tuning, we freeze the generator encoder and update its decoder.

Response to Q2

We separated CRM into two components, reconstruction and CLIP-guided semantic alignment, and extended the original ablation study with additional experiments.

Table 3: Ablation study on the GenImage dataset. Results are reported as detection accuracy. ✓ and × indicate that a component is enabled or disabled, respectively. When all three components are disabled, evaluation is performed on the original fake images. The CRM module is enabled when both Reconstruction and Semantic Alignment are marked ✓.

Reconstruction OnlySemantic AlignmentASMUniFDDGS-NetEffortCLIP+LinearCLIP+LoRACNNSpotGramNetResNet50SAFENPRAvg.
✘✘✘0.700.610.870.640.690.430.540.330.700.690.62
✔✘✘0.890.740.920.780.940.610.850.630.940.990.82
✔✔✘0.410.340.540.200.120.330.320.000.290.380.33
✔✘✔0.750.980.870.380.640.210.440.440.920.920.65
✘✘✔0.640.580.770.400.540.160.380.310.600.670.50
✔✔✔0.330.330.220.050.030.080.220.000.200.370.18

Analysis: Reconstruction without semantic optimization makes fake images easier to detect, while adding semantic alignment substantially improves attack effectiveness.

Reviewer_AmB7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors ask whether AI-generated image detectors pick up intrinsic evidence of forgery or incidental byproducts of generation. They group the cues detectors use into two families: semantic representations from vision foundation models, and synthesis-specific artifacts. They then build CRAS, a detector-agnostic anti-forensic framework with two cascaded modules. The Cross-Modal Reconstruction Module inverts a fake image into the Stable Diffusion latent space via DDIM inversion, then optimizes that latent with a pixel MSE term plus an alignment loss that pulls the CLIP image embedding toward a pool of 38 real-related text anchors and away from 2 fake-related anchors. The Artifact Suppression Module is a U-Net trained with a PatchGAN on content-matched pairs to remove the artifacts SD1.4 leaves behind. Validation covers 5 CLIP-based and 5 conventional detectors on GenImage, AIGCDetectBench, AIGIBench, and AIGI-Holmes, against 5 anti-forensic baselines. Average detection accuracy on GenImage falls from 0.62 on clean images to 0.18 under attack, at PSNR around 32.

Critical strengths

1.The module split turns the attack into a set of controlled experiments on the detectors. CRM handles semantics, ASM handles artifacts. Steering semantics alone drops average accuracy from 0.62 to 0.33; removing artifacts alone drops it to 0.50; doing both drops it to 0.18. The combined drop of 0.44 sits close to the sum of the single drops, 0.41, so the two evidence channels carry complementary information.

2.The attack preserves visual quality without touching the detectors. Latent-space optimization through a frozen SD1.4 produces adversarial images with no visible noise, while DiffAttack, FTM, and FPR all leave visible perturbations in Fig. 3. The method requires no detector weights, outputs, or gradients, only the CLIP encoders and SD1.4.

3.The evaluation is broad, and the cross-generator claim comes with a clear mechanism. Inference first passes every input, whatever its source generator, through the same fixed SD1.4 reconstruction, so the ASM trained on DRCT pairs faces exactly the artifacts it learned to remove. The diagnostic analyses tie each module to its intended effect.

Critical weaknesses
  1. Every adversarial image is generated by SD1.4 reconstruction followed by a trained post-processing network. Two plausible explanations exist for why detectors fail on such images: the attack steered and suppressed the semantic and artifact evidence, or the regeneration through SD1.4 plus ASM shifted the image away from the detectors' training distribution. Table 5 sets up no ablation entry that separates the two hypotheses. The data in Table 8 further suggests that most of the performance loss appears before semantic steering takes effect.

  2. Some detectors collapse under the attack yet performed near chance before it. Post-attack accuracy close to zero for these models does not demonstrate that the attack manipulated forensic evidence. The evaluation also omits two detector families central to the paper's argument: reconstruction-error methods (Dire) and frequency-domain forensics (Synthbuster), for which the ForensicHub benchmark (NeurIPS 2025) provides unified implementations.

  3. The CLIP2 experiments in the appendix enrich the coverage of vision foundation models but remain confined to the semantic paradigm. Since the adversarial images are SD reconstructions with spectra aligned toward real images, these two detector families are exactly suited to test the core claim of artifact suppression. The paper also lacks robustness results under JPEG compression and rescaling, as well as a false-positive evaluation on real images; both are necessary to support the threat model in deployment scenarios.

The extended ablation resolves my main concern. Reconstruction alone makes the fake images easier to detect. The drop under attack comes from semantic steering, not from regeneration shift. The added evaluations on reconstruction-based and frequency-domain detectors close the coverage gap. So I keep my score.

Questions for authors
  1. Evaluate at least one reconstruction-error detector and one frequency-domain detector; the ForensicHub implementations can serve. Do they catch the adversarial images?

  2. Do the adversarial images survive JPEG compression and resizing, and do they change detector scores on real images?

2 comments

We thank the reviewer for the positive assessment. We have conducted additional experiments to address your concerns.


Response to Weakness 1

We separated CRM into two components, reconstruction and CLIP-guided semantic alignment, and extended the original ablation study with additional experiments.

Table 1: Ablation study on the GenImage dataset. Results are reported as detection accuracy. ✓ and × indicate that a component is enabled or disabled, respectively. When all three components are disabled, evaluation is performed on the original fake images. The CRM module is enabled when both Reconstruction and Semantic Alignment are marked ✓.

Reconstruction OnlySemantic AlignmentASMUniFDDGS-NetEffortCLIP+LinearCLIP+LoRACNNSpotGramNetResNet50SAFENPRAvg.
✘✘✘0.700.610.870.640.690.430.540.330.700.690.62
✔✘✘0.890.740.920.780.940.610.850.630.940.990.82
✔✔✘0.410.340.540.200.120.330.320.000.290.380.33
✔✘✔0.750.980.870.380.640.210.440.440.920.920.65
✘✘✔0.640.580.770.400.540.160.380.310.600.670.50
✔✔✔0.330.330.220.050.030.080.220.000.200.370.18

Analysis:

  • The anti-forensic performance of the combination of SD1.4 reconstruction and ASM post-processing is inferior to that of the joint processing combination of CRM and ASM, which alters the semantics and suppresses artifacts simultaneously.
  • Reconstruction without semantic alignment makes the fake images easier to detect, indicating that detection performance drops only after semantic guidance is applied.

Response to Weakness 3 & Q2

  • To evaluate the robustness of the adversarial images, we applied JPEG compression at quality factors of 0.95 and 0.75 and resized the images to 1024 × 1024 and 224 × 224 pixels. The detection results for the post-processed adversarial images are as follows:

Table 2: Detection Results for Post-Processed Adversarial Images on GenImage (Accuracy).

UniFDDGS-NetEffortCLIP+LinearCLIP+LoRACNNSpotGramNetResNet50SAFENPRSynthbuster
No Post-Processing0.330.330.220.050.030.080.220.000.200.370.14
JPEG Compression (Q = 0.95)0.460.220.210.100.020.120.270.010.610.150.28
JPEG Compression (Q = 0.75)0.240.110.290.130.010.130.180.000.020.000.28
Resize (1024 ×1024)0.320.330.220.050.020.070.200.000.180.360.29
Resize (224 × 224)0.340.330.220.050.030.100.270.000.370.280.29

Analysis: The attack remains effective after the adversarial images undergo JPEG compression and resizing.

  • The detection results for real images after the adversarial attack are as follows:

Table 3: Detection Results for Adversarial Images Generated from Real Images on GenImage (Accuracy). Accclean\mathrm{Acc}_{\mathrm{clean}} denotes detection accuracy evaluated exclusively on the original real images.

UniFDCNNSpotGramNetDIRE
Accclean\mathrm{Acc}_{\mathrm{clean}}0.950.970.970.98
Accadv\mathrm{Acc}_{\mathrm{adv}}0.990.990.991.00
ΔAcc\Delta\mathrm{Acc}+0.04+0.02+0.02+0.02

Analysis: Since our attack aims to have all images classified as real, applying it to real images increases detection accuracy.

Response to Weakness 2 & Q1

  • The explanation for "Some detectors collapse under the attack yet performed near chance before it" is that we focus exclusively on anti-forensic attacks against clean fake images without considering real images. The near-chance accuracy observed for some detectors before the attack reflects their baseline performance on fake images alone. Evaluating them on both real and fake images would better reflect their overall detection performance.
  • We have extended our evaluation using the ForensicHub [1] benchmark to include reconstruction-error-based methods (DRCT and DIRE) and frequency-domain methods (Synthbuster and HiFiNet). The results are as follows:

Table 4: Attack Performance Against Reconstruction-Based Detectors and Frequency-domain Detectors.

DatasetMetricsDRCT/Conv-BDRCT/UnivFDDIRESynthbusterHifinet
GenImageAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.640.32−0.32\begin{array}{c}0.64 \\\\ 0.32 \\\\ -0.32\end{array}0.890.59−0.30\begin{array}{c}0.89 \\\\ 0.59 \\\\ -0.30\end{array}0.990.78−0.21\begin{array}{c}0.99 \\\\ 0.78 \\\\ -0.21\end{array}0.340.14−0.20\begin{array}{c}0.34 \\\\ 0.14 \\\\ -0.20\end{array}0.410.22−0.19\begin{array}{c}0.41 \\\\ 0.22 \\\\ -0.19\end{array}
AIGCDetectBenchAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.360.24−0.12\begin{array}{c}0.36 \\\\ 0.24 \\\\ -0.12\end{array}0.910.64−0.27\begin{array}{c}0.91 \\\\ 0.64 \\\\ -0.27\end{array}0.820.73−0.09\begin{array}{c}0.82 \\\\ 0.73 \\\\ -0.09\end{array}0.280.16−0.12\begin{array}{c}0.28 \\\\ 0.16 \\\\ -0.12\end{array}0.630.41−0.22\begin{array}{c}0.63 \\\\ 0.41 \\\\ -0.22\end{array}
AIGIBenchAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.440.17−0.27\begin{array}{c}0.44 \\\\ 0.17 \\\\ -0.27\end{array}0.880.70−0.18\begin{array}{c}0.88 \\\\ 0.70 \\\\ -0.18\end{array}0.940.81−0.13\begin{array}{c}0.94 \\\\ 0.81 \\\\ -0.13\end{array}0.230.10−0.13\begin{array}{c}0.23 \\\\ 0.10 \\\\ -0.13\end{array}0.360.13−0.23\begin{array}{c}0.36 \\\\ 0.13 \\\\ -0.23\end{array}
AIGI-HolmesAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.520.35−0.17\begin{array}{c}0.52 \\\\ 0.35 \\\\ -0.17\end{array}0.800.61−0.19\begin{array}{c}0.80 \\\\ 0.61 \\\\ -0.19\end{array}0.710.59−0.12\begin{array}{c}0.71 \\\\ 0.59 \\\\ -0.12\end{array}0.120.08−0.04\begin{array}{c}0.12 \\\\ 0.08 \\\\ -0.04\end{array}0.230.15−0.08\begin{array}{c}0.23 \\\\ 0.15 \\\\ -0.08\end{array}
Avg.AcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.490.27−0.22\begin{array}{c}0.49 \\\\ 0.27 \\\\ -0.22\end{array}0.870.63−0.24\begin{array}{c}0.87 \\\\ 0.63 \\\\ -0.24\end{array}0.860.72−0.14\begin{array}{c}0.86 \\\\ 0.72 \\\\ -0.14\end{array}0.240.12−0.12\begin{array}{c}0.24 \\\\ 0.12 \\\\ -0.12\end{array}0.400.22−0.18\begin{array}{c}0.40 \\\\ 0.22 \\\\ -0.18\end{array}

Analysis: The detection performance of both reconstruction-based and frequency-domain methods declines, demonstrating that our method is effective against both types of detectors.

References:[1] Du B, Zhu X, Ma X, et al. ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization[C]//Advances in Neural Information Processing Systems. 2025, 38: 72073–72102.

Reviewer_B6KbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, we focus on whether the existing AIGI Detector is vulnerable to dependence on Semantic Cues and Artifact Cues. It also attempts to reveal the potential Representation Bias of the detector by Black-box Anti-Forensic Attack.

The authors propose CRAS, which implements the attack through two cascaded modules: CRM leverages CLIP-guided Diffusion Reconstruction to align the visual representations of the generated images to the true semantic anchors; ASM constructs paired data using real images and their diffused reconstructions to learn to suppress the generative artifacts introduced during reconstruction. Experiments covering four public datasets and ten detectors show that the proposed method can effectively reduce the detection accuracy while maintaining high image quality.

Critical strengths
  1. This paper investigates the reliability of AIGI Detector from an Anti-Forensic perspective, focusing on whether existing detectors over-rely on semantic bias and specific artifacts in the generation process. This problem helps to understand the potential vulnerability of detectors.

  2. CRAS do not require access to the parameters or gradients of the target Detector, are able to attack both CLIP-based and Conventional Detectors, and exhibit good cross-detector migration capabilities.

Critical weaknesses
  1. The proposed framework mainly combines existing CLIP-guided latent optimization and image restoration techniques. CRM relies on conventional diffusion reconstruction and cross-modal feature alignment, while ASM adopts a standard encoder-decoder architecture with GAN and L1 losses for paired image restoration. The overall framework primarily consists of a cascade of established techniques, without clearly demonstrating the methodological novelty.

  2. The paper attributes the attack effectiveness to the vulnerabilities of semantic and artifact cues. However, the current experiments do not sufficiently disentangle these effects from the distribution shifts and low-level statistical changes introduced by diffusion reconstruction itself. In particular, ASM primarily learns to suppress SD reconstruction artifacts rather than artifacts originating from the input generators. Therefore, the experimental evidence does not fully support the claim that the attack effectiveness stems from independently manipulating these two types of forensic evidence.

  3. Limited Evaluation on Reconstruction-based Detectors. Although the paper evaluates various CLIP-based and conventional detectors, it lacks evaluations on recent reconstruction-based detection approaches, such as DDA. Given that CRAS heavily relies on diffusion reconstruction, evaluating such detectors is particularly important for assessing the generalizability and effectiveness of the proposed attack.

[1] Chen J, Chen H, Chen K, et al. Diffusion models for imperceptible and transferable adversarial attack[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 47(2): 961-977.

[2] Isola P, Zhu J Y, Zhou T, et al. Image-to-image translation with conditional adversarial networks[C]//2017 IEEE conference on computer vision and pattern recognition (CVPR). Ieee, 2017: 5967-5976.

Questions for authors
  1. What are the core technical differences between CRM and the existing CLIP-guided Diffusion Optimization Method, and ASM and the traditional Paired Image Restoration Method?

  2. How do the authors distinguish the effects of semantic manipulation and artifact suppression from the distribution shifts introduced by diffusion reconstruction itself? The current ablations demonstrate the effectiveness of CRM and ASM, but do not establish that they independently target the claimed forensic evidence. A reconstruction-only baseline and more direct analyses of semantic and artifact changes would help validate the proposed mechanisms.

  3. How does CRAS perform against recent reconstruction-based AIGI detectors, such as DDA? Given that ASM is trained to suppress artifacts introduced by stable diffusion reconstruction, it remains unclear whether its effectiveness generalizes to detectors trained with reconstruction-based data alignment strategies. Additional evaluations would help establish the generality of the proposed attack.

[1] Chen B, Zeng J, Yang J, et al. DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection of Diffusion Generated Images[C]//ICML. 2024, 2(4): 6.

[2] Chen R, Xi J, Yan Z, et al. Dual data alignment makes ai-generated image detector easier generalizable[J]. arXiv preprint arXiv:2505.14359, 2025.

2 comments

We thank the reviewers for their constructive feedback. Below, we clarify how CRM and ASM differ technically from existing methods and present additional experiments to isolate the effect of diffusion reconstruction itself and evaluate the attack effectiveness of CRAS against reconstruction-based detectors.


Response to Q1 & Weakness 1

We explain how CRM and ASM each differ from existing methods as follows:

  • Both CRM and DiffAttack [1] use DDIM inversion and optimize a diffusion latent for each image. DiffAttack generates adversarial examples against a surrogate classifier and relies on those examples to transfer to other classifiers. CRM targets CLIP, an upstream feature extractor used by AI-generated image detectors, enabling the attack to transfer across detectors without accessing the detectors themselves. Compared with existing CLIP-guided diffusion optimization methods, we designed two text pools containing the most representative real and fake prompts. These prompts achieve excellent performance in zero-shot detection with CLIP. We also designed a different alignment loss to make the CLIP embeddings of reconstructed images match the real prompts more strongly.
  • The technical differences between ASM and conventional paired image restoration methods such as pix2pix [2] lie in the training pairs and how the model is used. We train ASM to map each DRCT Stable Diffusion reconstruction to its corresponding original MS-COCO image to suppress artifacts left by reconstruction. To adapt the model to the anti-forensic task, we modify its convolutional layers. During fine-tuning, we freeze the generator encoder and update its decoder.

References:[1] Chen J, Chen H, Chen K, et al. Diffusion models for imperceptible and transferable adversarial attack[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 47(2): 961-977.

[2] Isola P, Zhu J Y, Zhou T, et al. Image-to-image translation with conditional adversarial networks[C]//2017 IEEE conference on computer vision and pattern recognition (CVPR). Ieee, 2017: 5967-5976.


Response to Q2 & Weakness 2

To investigate the effect of diffusion reconstruction itself, we separated CRM into two components, reconstruction and CLIP-guided semantic alignment, and extended the original ablation study with additional experiments. The results in Table 1 show that reconstruction alone makes the fake images easier to detect because it introduces new VAE-specific artifacts. The attack becomes effective only when semantic alignment is incorporated into the reconstruction optimization. Moreover, applying ASM after reconstructing the original fake images is less effective at evading detection than applying ASM directly to those images.

The reviewer notes that “ASM primarily learns to suppress SD reconstruction artifacts rather than artifacts originating from the input generators.” In the full pipeline, every input image undergoes the same fixed SD reconstruction process, regardless of its source generator. This reconstruction disrupts the input image’s original artifact patterns, and ASM learns to suppress the artifacts present in the CRM output.

Table 1: Ablation study on the GenImage dataset. Results are reported as detection accuracy. ✓ and × indicate that a component is enabled or disabled, respectively. When all three components are disabled, evaluation is performed on the original fake images. The CRM module is enabled when both Reconstruction and Semantic Alignment are marked ✓.

Reconstruction OnlySemantic AlignmentASMUniFDDGS-NetEffortCLIP+LinearCLIP+LoRACNNSpotGramNetResNet50SAFENPRAvg.
✘✘✘0.700.610.870.640.690.430.540.330.700.690.62
✔✘✘0.890.740.920.780.940.610.850.630.940.990.82
✔✔✘0.410.340.540.200.120.330.320.000.290.380.33
✔✘✔0.750.980.870.380.640.210.440.440.920.920.65
✘✘✔0.640.580.770.400.540.160.380.310.600.670.50
✔✔✔0.330.330.220.050.030.080.220.000.200.370.18

Response to Q3 & Weakness 3

We evaluated our attack against DDA and DRCT, two reconstruction-based detectors for AI-generated images. The results are shown below:

Table 2: CRAS Attack Performance Against Reconstruction-Based Detectors.

DatasetMetricsDRCT/Conv-BDRCT/UnivFDDDA
GenImageAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.640.32−0.32\begin{array}{c}0.64 \\\\ 0.32 \\\\ -0.32\end{array}0.890.59−0.30\begin{array}{c}0.89 \\\\ 0.59 \\\\ -0.30\end{array}0.800.67−0.13\begin{array}{c}0.80 \\\\ 0.67 \\\\ -0.13\end{array}
AIGCDetectBenchAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.360.24−0.12\begin{array}{c}0.36 \\\\ 0.24 \\\\ -0.12\end{array}0.910.64−0.27\begin{array}{c}0.91 \\\\ 0.64 \\\\ -0.27\end{array}0.830.75−0.08\begin{array}{c}0.83 \\\\ 0.75 \\\\ -0.08\end{array}
AIGIBenchAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.440.17−0.27\begin{array}{c}0.44 \\\\ 0.17 \\\\ -0.27\end{array}0.880.70−0.18\begin{array}{c}0.88 \\\\ 0.70 \\\\ -0.18\end{array}0.730.64−0.09\begin{array}{c}0.73 \\\\ 0.64 \\\\ -0.09\end{array}
AIGI-HolmesAcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.520.35−0.17\begin{array}{c}0.52 \\\\ 0.35 \\\\ -0.17\end{array}0.800.61−0.19\begin{array}{c}0.80 \\\\ 0.61 \\\\ -0.19\end{array}0.930.81−0.12\begin{array}{c}0.93 \\\\ 0.81 \\\\ -0.12\end{array}
Avg.AcccleanAccadv,↓ΔAcc,↓\begin{array}{l} \mathrm{Acc}_{\mathrm{clean}} \\\\ \mathrm{Acc}_{\mathrm{adv}}\\,\downarrow \\\\ \Delta\mathrm{Acc}\\,\downarrow \end{array}0.490.27−0.22\begin{array}{c}0.49 \\\\ 0.27 \\\\ -0.22\end{array}0.870.63−0.24\begin{array}{c}0.87 \\\\ 0.63 \\\\ -0.24\end{array}0.820.71−0.11\begin{array}{c}0.82 \\\\ 0.71 \\\\ -0.11\end{array}

Analysis: CRAS transfers to both DRCT variants, while its effect on DDA is modest because DDA aligns reconstructed images with real images in both the frequency and pixel domains.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.