acceptodds
Withdrawn from ICLR 2027

Frequency Attribution Mask: Toward Detector-Adaptive Robustness in AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image (AIGI) detectors report high accuracy on benchmark splits, yet collapse on unseen generators and after routine post-processing such as JPEG compression or resizing, exposing a reliance on shortcut cues that do not transfer. Existing remedies name a specific source-level bias and patch it with a deterministic operation at the data, architecture, or weight level, leaving every other bias untouched and tying the fix to a single detector setup. To address this, we propose *Frequency Attribution Mask* (FAM), an attribution-guided mask-tuning framework that intervenes at each detector's own input sensitivity, the point at which any such bias actually manifests. FAM derives a per-sample frequency-domain attribution from the detector's input gradient, masks the highest-attribution components of the input spectrum with class-asymmetric thresholds, and finetunes on the masked reconstructions. We apply FAM to two pretrained ResNet-50 detectors, one of which already carries a pixel-level alignment fix, and show that it consistently improves post-processing robustness over both base detectors and recent shortcut-mitigation baselines, with the largest gains where residual shortcut reliance is most severe.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Image Detection, Frequency Attribution, Shortcut Mitigation, Post-Processing Robustness

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jyJ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on the performance degradation of AI-generated image detectors under common post-processing operations and when the generator distribution is shifted. It points out that the common cause might be that the model relies on shortcut features that are not generalizable. In response to the limitations of existing methods, which typically intervene in specific data biases, the authors propose Frequency Attribution Mask (FAM), which serves as a post-hoc remedy for shortcut learning. FAM constructs a frequency attribution map using the input gradients of the detector, masks the high-attribution frequency components with a class asymmetry threshold, and then fine-tunes the masked reconstructed images. Experiments on two pre-trained detectors, Corvi and Rajan, show that this method improves the detection robustness of the model under unseen generators and post-processing conditions, and achieves better results than the original detector on real-world benchmark scenarios.

Critical strengths
  1. This paper addresses a relevant and practical problem. Improving the robustness of AIGI detectors to common post-processing operations and unseen generators is important for real-world applications.
  2. This paper provides a clear motivation for the proposed method. The authors systematically analyze the limitations of existing approaches and propose a detector-adaptive method that identifies frequency cues based on the detector’s own input gradients.
  3. Experiments on two pretrained detectors show improvements under post-processing and unseen-generator settings, with additional evaluations on in-the-wild benchmarks.
Critical weaknesses
  1. Gradient sensitivity is not established as a proxy for shortcut reliance. Highly attributed frequencies may encode transferable cues rather than shortcuts. FAM could therefore remove useful information, while fine-tuning on masked inputs may introduce new shortcuts.
  2. The contribution of frequency attribution remains unclear. In Table 2, random frequency masking yields substantial gains, while spatial attribution masking performs similarly to FAM. These results do not fully isolate attribution-specific benefits from masking and fine-tuning.
  3. The paper does not report the fraction of frequency components removed or the spectral energy retained, leaving sample-wise intervention levels and their effects unclear.
  4. Direct comparisons of real and fake frequency attribution maps are needed to support the claimed band differences and motivate asymmetric rather than shared thresholds.
  5. Evaluating only two pretrained detectors leaves uncertainty about how training stage and initial performance affect FAM’s effectiveness.
Questions for authors
  1. What evidence supports linking highly attributed frequencies to non-robust shortcuts rather than transferable cues? Does masked fine-tuning introduce new shortcuts?
  2. Can FAM, random frequency masking, spatial attribution masking, and the unmasked baseline be compared under matched training settings and, where applicable, masking strengths to isolate the sources of gains?
  3. Can the authors report class-wise masking ratios (mean and variance) and retained spectral energy?
  4. Can attribution visualizations and quantitative analysis substantiate real–fake frequency differences, alongside a comparison of class-asymmetric and shared thresholds?
  5. How does FAM perform across early, intermediate, and converged checkpoints of the same detector?
Reviewer_Vyg4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Frequency Attribution Mask (FAM), a finetuning framework for improving AI-generated image detection under unseen generators and common post-processing. Its main contributions are: 1) deriving per-image frequency attribution maps from the input gradients of a pretrained detector to identify its sensitive frequency components; 2) applying class-asymmetric masks to suppress high-attribution components and finetuning the detector on the reconstructed images without changing its architecture; and 3) demonstrating improvements over two ResNet-50 base detectors on post-processed and in-the-wild benchmarks, with ablations examining the roles of frequency-domain masking and attribution-guided selection.

Critical strengths
  • FAM uses a detector’s own sensitivity to construct training perturbations without modifying its architecture or adding an attribution stage at inference. This makes the approach potentially useful for improving existing detectors.
  • Table 1 reports improvements in both AUC and balanced accuracy over both base detectors, including the already aligned Rajan model. The in-the-wild evaluations broaden the evidence, while the unmasked-finetuning and random-mask controls help establish that the gains are not explained solely by additional training.
Critical weaknesses
  • The link between high attribution and shortcut reliance needs clarification. Sections 2.1 and 3.1 motivate suppressing high-attribution frequencies as a remedy for shortcut reliance. However, high sensitivity does not by itself distinguish unstable shortcuts from useful, transferable generative traces. Could the authors clarify whether this correspondence is supported by direct evidence or is a working hypothesis? If evidence is available, making it explicit would strengthen the motivation; otherwise, targeted validation and a clearer statement of the hypothesis would be helpful. This distinction matters to whether FAM can be interpreted as identifying shortcuts rather than important features more generally.
  • The correspondence between the attribution score and the masking operation remains insufficiently established. Equation (3) measures local sensitivity through gradient magnitude, whereas Equations (4)–(5) remove coefficients entirely. The effect of deletion also depends on coefficient magnitude, phase, and model nonlinearity. Could the authors explain why gradient magnitude alone is an appropriate selection criterion for this operation? The MoRF analysis provides useful evidence, but further validation on the initial detector that supplies the training masks would help substantiate the selection rule.
  • The explanation for improved generalization appears stronger than the current mechanistic evidence. The performance gains and attribution shifts support the effectiveness of FAM, but do not yet clearly distinguish shortcut removal and more stable feature learning from general augmentation or regularization effects. In particular, concentrated frequency dependence does not by itself imply robustness. Could the authors provide stronger evidence connecting the learned frequency support to its stability under post-processing and generator shifts? If this explanation remains a hypothesis, explicitly identifying it as such and providing focused supporting evidence would make the conclusions easier to assess.
Questions for authors
  • Could the authors identify evidence that directly connects the selected frequencies to unstable cues? In particular, how well does the ranking predict deletion effects in the initial detector, and what evidence distinguishes shortcut suppression from general masking benefits? A controlled intervention on a known spurious cue would be one possible way to examine the latter.
  • Does the claim in Sections 2.2–3.1 that spatial saliency cannot isolate these cues refer specifically to the pixelwise thresholding baseline? A globally distributed cue can still be represented in the spatial gradient. In Table 2, Rajan’s spatial-attribution baseline and FAM both reach 87.8 balanced accuracy, with AUCs of 94.1 and 95.2. Could the authors clarify how masking strength and tuning opportunities were controlled, and report variability across repeated runs for this key comparison? Appendix B.5 already matches random-mask sparsity; the remaining question concerns the spatial-attribution comparison and the reliability of the incremental gain.
  • How were decision thresholds determined for each method, and were comparable validation-based calibration procedures used? Table 1 gives Corvi+ higher AUC than CorviF (89.5 versus 87.4), but lower balanced accuracy (73.9 versus 81.9). Could the authors clarify this tradeoff and provide classwise accuracies or detection rates at a fixed false-positive rate to help interpret the practical benefit?
  • Lines 299–300 describe applying post-processing to synthetic images, whereas Appendix C.3 refers to each test image. Were both real and generated images processed? Were all methods evaluated on the same transformed samples, with the same sampled parameters and operation order? Clarifying this would help establish precisely which robustness setting Table 1 measures.
  • Could the authors provide paired clean and post-processed results for the full nine-generator evaluation in Table 1? Figure 1 and the unperturbed MoRF reference already provide some clean-input results, but a complete paired comparison would clarify whether the robustness gains involve a tradeoff in clean-input detection.
Reviewer_5HgcOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper deals with the challenge of improving the robustness of AIGI detectors. The authors propose a method that augments the training data with images whose frequency band of the highest sensitivity to the base detector is masked. Experimental results show that the proposed method can improve two ResNet-50 detectors’ robustness against multiple post-processing methods.

Critical strengths

I appreciate the authors’ attempt to improve AIGI detection robustness in a higher level than previous work. Experiments show that the proposed method is effective under multiple post-processing methods including JPEG and resizing. The paper is easy to understand with a straightforward illustration.

Critical weaknesses

I think the depth of this paper is limited and I don’t learn much from it as a researcher who is interested in robustness but not particularly AIGI detection. The idea of obtaining robustness through data augmentation in the frequency domain is not new (e.g., “A Fourier-Based Framework for Domain Generalization. CVPR 2021.”) and the idea of using data augmentation to avoid reliance of shortcut/non-robust features is also well-established (“Adversarial Examples Are Not Bugs, They Are Features.”).

The evidence provided by the authors hardly support claims beyond that FAM is a working trick that improves AIGI detection robustness. For example, in Line 439, the authors claim “FAM instead locates the offending cue” and in 442 “removing the bands a detector leans on rehearses the loss post-processing will cause”, but I don’t see any analysis showing that the affect caused by the post-processing can be located by FAM in the frequency domain.

The paper claims that FAM is a unified remedy for shortcut learning (L473) while there could be many shortcuts that cannot be tested by the post-processing methods used in the experiments, such as semantic, subject, composition, style, watermark, camera, etc. I think the authors should define the applicable scope of the proposed method more responsibly instead of claiming it is a unified method simply because it’s effective against more than one shortcut.

Questions for authors

Do JPEG, resizing, WebP, and color jittering happen to only affect the bands where will be masked by FAM in the frequency domain? If not, how do you support the claims that “FAM locates the offending cue” and “removing the band rehearses the loss post-processing will cause”?

Why only ResNet50 is used in the experiment? Does FAM also improves ViT-based detectors?

How many fake images are used in the experiments? How reliable are the quantitative results statistically?

Why do you omit WebP in Figure 1 and omit Color Jitter in Figure 4?

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

Related papers

Discussion (0)

Sign in to comment.