acceptodds
Under review as a conference paper at ICLR 2027

Carrier-Perturbation Responses: An Alternative Basis for Generalizable AI-Generated Image Detection

Abstract

Existing generated image detectors typically infer image authenticity from features extracted from the image itself. However, such features may vary substantially across scenes, limiting generalization from unseen domains. Our preliminary observations show that, under the same controlled perturbations, real and generated images exhibit different feature-change distributions in the representation space of a frozen visual foundation model, while responses from same-class images remain relatively consistent across scenes. This motivates us to investigate perturbation-induced feature changes as an alternative basis for authenticity assessment. To this end, we propose Carrier-Perturbation Responses (CPR), which treats each image as a carrier and characterizes its response to a controlled perturbation through the feature change between the original and perturbed images. Since some perturbations may produce responses dominated by the perturbation rule itself or responses too weak for discrimination, we further develop an ANOVA-inspired selection criterion that favors perturbations with strong carrier–perturbation interactions, limited rule-induced effects, and sufficient response magnitude. Together, these designs shift the detection basis from scene-dependent appearance features toward perturbation-induced response patterns, providing complementary evidence for authenticity assessment across unseen scenes. Across four in-the-wild datasets, CPR improves the average real-image accuracy by 3.3 percentage points to 96.7% and the average overall accuracy by 2.1 percentage points to 95.4%. Our code is available at https://anonymous.4open.science/r/CPR-F211.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Carrier-Perturbation Response, image forensics, domain generalization, image authenticity

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NGsFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Carrier-Perturbation Responses (CPR) for AI-generated image detection. Instead of classifying an image from its own features, CPR applies a controlled perturbation to the image and uses the resulting feature change, Φ(x + δ) − Φ(x), from a frozen DINOv3-7B encoder as the representation. The idea is that real and generated images respond differently to the same perturbation, and that these responses depend less on scene content than the original features do. To pick perturbations, the authors propose an ANOVA-inspired criterion that splits response variation into carrier, perturbation and interaction effects. It favors perturbations with strong interaction effects, weak rule-induced effects and sufficient response strength. One perturbation is selected from each of three groups (frequency, local structure, attention-guided), a linear head is trained on each, and their predictions are averaged. Experiments on GenImage and four in-the-wild datasets show improvements over the DINOv3-Linear baseline, from 96.4% to 97.8% average accuracy on GenImage and from 93.3% to 95.4% on the in-the-wild datasets, along with robustness to JPEG compression and blur.

Critical strengths
  • Simple, practical idea with a clear ablation. CPR adds no new trainable modules: it uses a frozen encoder, a fixed perturbation and a linear head, so it is cheap to train and easy to reproduce. Table 6 directly compares Φ(x), Φ(x + δ), Φ(δ) and the difference Φ(x + δ) − Φ(x), and shows that the feature change outperforms the other choices. This gives reasonable support for the core design decision.
  • Broad evaluation on in-the-wild data. Beyond GenImage, the method is tested on four in-the-wild datasets (Chameleon, WildRF, SocialRF, SynthWildX) and under JPEG compression and blur. CPR improves over the same-backbone DINOv3-Linear baseline on all four datasets, with the clearest gains in real-image accuracy. This is the failure mode where many prior detectors break down, as the very low real or fake accuracies of several baselines in Table 4 show.
Critical weaknesses
  • Backbone dependence. Most of CPR's improvement over prior detectors seems to come from the DINOv3-7B backbone rather than from CPR itself. The same-backbone DINOv3-Linear baseline already reaches 96.4% on GenImage and 93.3% in the wild, and CPR adds only 1.4 and 2.1 points on top of that. The authors should apply CPR to other backbones, such as DINOv2 (used by DDA) or CLIP (used by UnivFD), to allow a fair comparison with these methods.
Questions for authors

Please address the points raised in the Weaknesses section (I would be willing to raise my score.)

1 comment
AuthorsOpenReview ↗

Evaluating CPR Beyond DINOv3

We thank the reviewer for recognizing the simplicity of CPR, the representation ablation, and the evaluation on in-the-wild datasets. We agree that the contribution of CPR should be distinguished from the performance advantage provided by the pretrained backbone, and that evaluation beyond DINOv3 is important. To address this concern, we additionally evaluated CPR with PE-CLIP, SigLIP2, CLIP, and DINOv2. These additional experiments show that the benefits of CPR extend beyond DINOv3, while also revealing differences across backbones and evaluation domains.

Results without CPR. All values are percentages.

DatasetPE-CLIP + LinearSigLIP2 + LinearUnivFD (CLIP)DDA (DINOv2)DINOv2 + Linear
GenImage93.894.581.789.080.6
Chameleon89.982.260.282.468.0

Results with CPR. All values are percentages.

DatasetPE-CLIP + CPRSigLIP2 + CPRCLIP + CPRDINOv2 + CPR
GenImage95.194.981.189.8
Chameleon91.186.273.084.4

For PE-CLIP, CPR improves performance from 93.8% to 95.1% on GenImage and from 89.9% to 91.1% on Chameleon, corresponding to gains of 1.3 and 1.2 percentage points, respectively. For SigLIP2, the corresponding results improve from 94.5% to 94.9% and from 82.2% to 86.2%, yielding gains of 0.4 and 4.0 percentage points. These results support the effectiveness of perturbation-induced feature changes with backbones beyond DINOv3.

Following the reviewer's suggestion, we also evaluated CPR using the backbone families employed by DDA and UnivFD. DINOv2 + CPR achieves 89.8% on GenImage and 84.4% on Chameleon, compared with 89.0% and 82.4% for DDA, respectively. With CLIP, CPR achieves 73.0% on Chameleon, improving over UnivFD's 60.2% by 12.8 percentage points.

We agree that DINOv3 contributes substantially to the strong absolute performance of our detector. Our contribution is the use of perturbation-induced feature changes as an alternative representation for authenticity assessment, together with a criterion for selecting effective perturbations. The additional experiments support the usefulness of this design beyond DINOv3, while also showing that its benefits depend on the backbone and evaluation domain. We thank the reviewer for encouraging a more direct experimental examination of this distinction.

Reviewer_gceqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Carrier-Perturbation Responses (CPR) for AI-generated image detection. Instead of classifying an image directly from its original representation, CPR applies controlled perturbations and uses the resulting feature differences, extracted with a frozen DINOv3-7B encoder, as authenticity cues. An ANOVA-inspired criterion decomposes response variation into carrier, perturbation-rule, and interaction components, and combines these with response magnitude to select one perturbation from each of three candidate groups. Separate linear classifiers are trained on the selected responses, and their predicted probabilities are averaged. The experiments cover GenImage and four in-the-wild datasets.

Critical strengths
  1. An interesting alternative detection signal. Characterizing how an image's representation changes under a controlled perturbation, rather than relying exclusively on its unperturbed features, is an intuitive and potentially useful direction. Freezing the foundation encoder and training only linear heads also helps isolate the discriminative value of these response features.

  2. Relevant comparisons and ablations. The paper includes a DINOv3-based linear baseline, compares original, perturbed, and difference features, and evaluates individual perturbations and their ensemble. These comparisons provide useful evidence that the improvement is not solely attributable to the strong pretrained encoder.

Critical weaknesses
  1. Numerical inconsistencies in the reported accuracies. Table 4 defines overall ACC as the sample-proportion-weighted accuracy over real and fake samples. However, on SynthWildX, DINOv3-Linear has Real = 0.917, Fake = 0.913, and ACC = 0.909. Under the stated definition, overall ACC must lie between the two class-wise accuracies, so this triplet is mathematically inconsistent and cannot be explained by rounding. There are further discrepancies for WildRF, which Table 7 lists as having 2,650 real and 2,650 fake images. For example, Gram-Net's Real = 0.947 and Fake = 0.205 imply ACC ≈ 0.576 on that balanced dataset, rather than the reported 0.543; similar discrepancies occur for NPR, SAFE, Effort, and AIDE. These issues raise concerns about the evaluation or aggregation procedure. The authors should provide exact evaluation counts and confusion matrices, correct the affected entries, and verify whether the reported averages and improvements change after correction.

  2. Insufficient evidence that the proposed selection score generalizes beyond the evaluated candidates. The ANOVA-inspired variance decomposition is a reasonable descriptive tool, but a large carrier–perturbation interaction does not, by itself, establish that a response contains more information about the real/fake label. The results support the score's usefulness within the eleven predefined candidates, and Appendix D analyzes alternative weight configurations. However, that analysis reuses the same candidates and their measured accuracies; it does not test whether the criterion reliably predicts discriminative performance for new perturbations or a different training distribution. A stronger validation would fix the scoring weights in advance, evaluate an independently constructed set of perturbations, and compare the selected candidates against simpler rules, such as selecting by response magnitude alone or by a single variance component.

Questions for authors
  1. Can the authors reconcile all Real/Fake/ACC triplets in Table 4 with the dataset compositions reported in Table 7? Please provide exact class counts, confusion matrices, and the aggregation procedure. If macro- and micro-averaged metrics were mixed, please identify them explicitly and recompute the main comparisons consistently.
  2. How were the selection weights and perturbation hyperparameters determined, and which data splits were used? Were these choices finalized before examining the reported test accuracies? Can the authors demonstrate that the fixed selection criterion remains effective on previously unseen perturbation candidates or a separate selection dataset?
  3. What numerical values of the carrier count, perturbation-rule count, and number of repeated experiments were used in the variance decomposition? Please specify the operational definitions of all eleven candidates, including those not selected. How does CPR compare with an ordinary multi-view ensemble under a matched encoder-forward-pass budget, and what are its practical inference latency and peak memory requirements?

I would be willing to reconsider my assessment if these issues are adequately resolved.

5 comments
AuthorsOpenReview ↗

Corrections and Complete-Data Reevaluation

We thank the reviewer for identifying these inconsistencies and apologize for our mistakes. In the original Table 4, we accidentally swapped the SynthWildX Fake accuracy and ACC for DINOv3-Linear. The corrected values for that run are Real = 0.917, Fake = 0.909, and ACC = 0.913.

Table 7 reported the official dataset size, but we failed to check the files actually used. Some downloaded files were corrupted or missing; the WildRF split in the earlier run contained 1,044 real and 1,250 generated images. We recalculated the weighted ACC values using the actual counts. The corrected earlier-run results are shown in Table A.

We later obtained the complete datasets and reran the evaluation. WildRF contains 1,250 images per class and SynthWildX contains 500 real and 1,500 generated images. Chameleon and SocialRF are unchanged. Table B updates only WildRF and SynthWildX, and recomputes the averages across all four datasets. For each dataset,

ACC=NrAccr+NfAccfNr+Nf.\mathrm{ACC}=\frac{N_r\mathrm{Acc}_r+N_f\mathrm{Acc}_f}{N_r+N_f}.

Avg. is the arithmetic mean across datasets, not a pooled accuracy.

Table A. Corrected earlier-run results. WildRF used 1,044 real and 1,250 generated images. Avg. is across four datasets. Each cell reports Real / Fake / ACC.

MethodChameleonWildRFSocialRFSynthWildXAvg.
Gram-Net0.992/0.044/0.5900.947/0.205/0.5430.531/0.523/0.5270.845/0.257/0.5450.829/0.257/0.551
NPR0.999/0.046/0.5940.980/0.243/0.5780.593/0.537/0.5650.683/0.324/0.4990.814/0.288/0.559
UnivFD0.763/0.441/0.6260.693/0.637/0.6630.563/0.637/0.6000.545/0.509/0.5270.641/0.556/0.604
SAFE0.993/0.046/0.5900.918/0.309/0.5860.564/0.532/0.5480.772/0.283/0.5220.812/0.293/0.562
Effort0.394/0.782/0.5590.179/0.955/0.6020.513/0.533/0.5230.345/0.889/0.6220.358/0.790/0.577
DDA0.940/0.708/0.8410.899/0.908/0.9040.818/0.846/0.8320.936/0.894/0.9150.898/0.839/0.873
OMAT0.899/0.359/0.6700.633/0.715/0.6780.581/0.633/0.6070.619/0.580/0.5990.683/0.572/0.639
AIDE0.944/0.203/0.6290.973/0.195/0.5490.578/0.541/0.5600.690/0.297/0.4890.796/0.309/0.557
DINOv3-Linear0.933/0.895/0.9170.948/0.975/0.9630.937/0.948/0.9420.917/0.909/0.9130.934/0.932/0.934
CPR(ρF,3\rho_{F,3})0.955/0.889/0.9270.976/0.976/0.9760.918/0.974/0.9460.949/0.932/0.9400.950/0.943/0.947
CPR(R∗R^{\ast})0.977/0.869/0.9310.987/0.976/0.9810.944/0.979/0.9620.958/0.935/0.9460.967/0.940/0.955

Table B. Complete-data rerun. WildRF: 1,250 real and 1,250 generated; SynthWildX: 500 real and 1,500 generated. Chameleon and SocialRF are unchanged. Avg. is across all four datasets. Each cell reports Real / Fake / ACC.

MethodChameleonWildRFSocialRFSynthWildXAvg.
Gram-Net0.992/0.044/0.5900.947/0.205/0.5760.531/0.523/0.5270.844/0.264/0.4090.829/0.259/0.526
NPR0.999/0.046/0.5940.982/0.243/0.6130.593/0.537/0.5650.682/0.330/0.4180.814/0.289/0.548
UnivFD0.763/0.441/0.6260.692/0.637/0.6640.563/0.637/0.6000.546/0.507/0.5170.641/0.556/0.602
SAFE0.993/0.046/0.5900.918/0.309/0.6140.564/0.532/0.5480.772/0.282/0.4050.812/0.292/0.539
Effort0.394/0.782/0.5590.175/0.955/0.5650.513/0.533/0.5230.344/0.889/0.7530.357/0.790/0.600
DDA0.940/0.708/0.8410.901/0.908/0.9040.818/0.846/0.8320.936/0.894/0.9050.899/0.839/0.871
OMAT0.899/0.359/0.6700.632/0.715/0.6740.581/0.633/0.6070.618/0.577/0.5870.683/0.571/0.635
AIDE0.944/0.203/0.6290.970/0.195/0.5830.578/0.541/0.5600.690/0.299/0.3970.796/0.310/0.542
DINOv3-Linear0.933/0.895/0.9170.946/0.975/0.9600.937/0.948/0.9420.918/0.905/0.9080.934/0.931/0.932
CPR(ρF,3\rho_{F,3})0.955/0.889/0.9270.973/0.976/0.9740.918/0.974/0.9460.948/0.933/0.9370.949/0.943/0.946
CPR(R∗R^{\ast})0.977/0.869/0.9310.989/0.976/0.9820.944/0.979/0.9620.958/0.934/0.9400.967/0.940/0.954
AuthorsOpenReview ↗

Probe Selection Protocol and Weight Validation

We thank the reviewer for raising this concern. We clarify that all statistics used to select the perturbations were computed exclusively on the SD-v1.4 training split. For each candidate perturbation, the statistics were computed separately for real and generated images and then averaged with equal class weights. None of the nine generator test subsets or the four in-the-wild datasets was used for selection.

The weights were specified a priori to reflect the intended priority among the statistics, rather than fitted to evaluation results. The interaction effect receives the largest positive weight because it captures the component expected to transfer across settings. The carrier effect receives a smaller positive weight, whereas the perturbation effect receives a negative weight because it reflects rule variation unrelated to the real-versus-generated distinction. Appendix D evaluates all six permutations of 3,1,−1\\{3,1,-1\\} and multiple weight spacings; (wint,wcar,wper)=(3,1,−1)(w_{\mathrm{int}},w_{\mathrm{car}},w_{\mathrm{per}})=(3,1,-1) achieves the best ranking agreement in all three groups. Appendix C.4 further analyzes the sensitivity to the instantiation parameters of the three perturbation types. Thus, the selection procedure and weights were fixed before evaluation on the test subsets and in-the-wild datasets.

As an additional validation, we applied the same scoring rule to the OpenImages results:

S(ρ)=3ηint+ηcar−ηper+Aˉ.S(\rho)=3\eta_{\mathrm{int}}+\eta_{\mathrm{car}}-\eta_{\mathrm{per}}+\bar{A}.

The table below reports the statistics shown in the OpenImages results and the corresponding scores. Scores were computed from the displayed statistics.

OpenImages statistics and scores for candidate perturbations.

FamilyProbeηcar\eta_{\mathrm{car}}ηper\eta_{\mathrm{per}}ηint\eta_{\mathrm{int}}Aˉ\bar{A}S(ρ)S(\rho)
RFR_FρF,1\rho_{F,1}0.69770.02420.27810.33861.8464
RFR_FρF,2\rho_{F,2}0.48750.10030.41220.50992.1337
RFR_FρF,3\rho_{F,3}0.43900.18720.37381.00002.3732
RSR_SρS,1\rho_{S,1}0.40150.23010.36841.00002.2766
RSR_SρS,2\rho_{S,2}0.41800.16030.42170.91792.4407
RSR_SρS,3\rho_{S,3}0.38040.18530.43430.89432.3923
RSR_SρS,4\rho_{S,4}0.35880.05470.58650.80952.8731
RAR_AρA,1\rho_{A,1}0.65230.10990.23781.00002.2558
RAR_AρA,2\rho_{A,2}0.55260.08750.35990.90512.4499
RAR_AρA,3\rho_{A,3}0.40190.11590.48220.80072.5333
RAR_AρA,4\rho_{A,4}0.38730.08040.53230.91822.8220

The OpenImages results select ρF,3\rho_{F,3}, ρS,4\rho_{S,4}, and ρA,4\rho_{A,4}, respectively, matching the original selection in all three groups. This serves as an additional cross-dataset validation and was not used to determine the original weights or selections.

AuthorsOpenReview ↗

Implementation Details and Computational Cost

(1) Scale of the variance decomposition.

The decomposition uses G=10G=10 replicate groups, run independently for the real and fake classes. Within each group and class, N=12N=12 carriers are drawn uniformly from the SD-v1.4 training split. For each candidate ρk\rho_k, N=12N=12 rules are sampled over its admissible range, with the primary parameter from U[0.2,0.4]\mathcal{U}[0.2,0.4]. Each group and class thus contributes N2=144N^2=144 CPRs. Across the K=11K=11 candidates, the selection stage evaluates G⋅2⋅N2⋅K=31,680G\cdot 2\cdot N^2\cdot K=31{,}680 CPRs. Each CPR requires one forward pass of the perturbed image. The carrier pass is computed once per group, class and candidate and shared across its NN rules, adding G⋅2⋅N⋅K=2,640G\cdot 2\cdot N\cdot K=2{,}640 passes. Selection therefore requires 34,32034{,}320 image-level forward passes in total (2,8602{,}860 batches of size NN). Statistics are pooled across groups before the ratios in Eq. (11), so ηq\eta_q is a ratio of summed sums of squares, not a mean of per-group ratios.

(2) Operational definitions of all eleven candidates.

Let xx denote the carrier, d(u,v)d(u,v) the normalized radial frequency, α\alpha the replacement proportion, pp the patch size (p=14p=14), and AclsA_{\mathrm{cls}} the head-averaged last-layer CLS attention of the frozen encoder.

Frequency group (RF\mathcal{R}_F):

  • ρF,1\rho_{F,1} is high-pass noise injection, x+ϵ,h~(n)x+\epsilon\\,\tilde{h}(n), where n∼N(0,I)n\sim\mathcal{N}(0,I) and ⋅~\tilde{\cdot} RMS-normalizes the high-pass residual h(⋅)h(\cdot) to ϵ\epsilon.
  • ρF,2\rho_{F,2} is high-pass component re-injection, x+ϵ,h~(x)x+\epsilon\\,\tilde{h}(x).
  • ρF,3\rho_{F,3} is frequency-band carrier filtering, F−1H(d)Fx\mathcal{F}^{-1}\\{H(d)\mathcal{F}x\\}, retaining only the prescribed band.

Local-structure group (RS\mathcal{R}_S). All candidates act on the edge-ranked patch set MM of fraction α\alpha:

  • ρS,1\rho_{S,1} is local inversion, x←1−xx\leftarrow 1-x on MM.
  • ρS,2\rho_{S,2} is neighborhood permutation, x←roll⁡(x,(p,p))x\leftarrow\operatorname{roll}(x,(p,p)) on MM.
  • ρS,3\rho_{S,3} disrupts edge continuity:

x←12roll⁡h(x)+12roll⁡w(x)−12sign⁡(∇2x)∣∇2x∣on M.x\leftarrow \tfrac{1}{2}\operatorname{roll}_{h}(x)+\tfrac{1}{2}\operatorname{roll}_{w}(x)-\tfrac{1}{2}\operatorname{sign}(\nabla^2x)|\nabla^2x|\quad\text{on }M.

  • ρS,4\rho_{S,4} is patch-wise noise injection, replacing a uniformly random fraction α\alpha of patches with Gaussian noise matching the image's mean and standard deviation.

Attention group (RA\mathcal{R}_A). All candidates act on the top-α\alpha attention patches MA=arg⁡max⁡AclsM_A=\arg\max A_{\mathrm{cls}}:

  • ρA,1\rho_{A,1} is masked noise injection, (1−s)(−x)+sN(1-s)(-x)+s\mathcal{N} on MAM_A.
  • ρA,2\rho_{A,2} exchanges MAM_A with the bottom-α\alpha attention patches.
  • ρA,3\rho_{A,3} is semantic smoothing, an α\alpha-blended mean blur on MAM_A.
  • ρA,4\rho_{A,4} is attention-guided masking, x←x⋅1[MˉA]x\leftarrow x\cdot\mathbf{1}[\bar{M}_A].

The selected candidates are ρF,3\rho_{F,3}, ρS,4\rho_{S,4} and ρA,4\rho_{A,4}.

(3) Forward-pass budget.

R∗R^{\ast} comprises the three selected perturbations, ρF,3\rho_{F,3}, ρS,4\rho_{S,4} and ρA,4\rho_{A,4}. It requires one shared carrier pass plus three perturbed-image passes, four encoder passes in total. Table 6 controls for pass count: the two-pass variants Φ(xi+δi)\Phi(x_i+\delta_i) and Φ(δi)\Phi(\delta_i) score 0.901/0.8660.901/0.866 and 0.885/0.8780.885/0.878, both below the one-pass Φ(xi)\Phi(x_i) baseline at 0.964/0.9330.964/0.933. Extra encoder evaluations alone thus do not explain the gain from 0.9640.964 to 0.974/0.9460.974/0.946; the gain comes from the differencing operation. This also separates CPR from multi-view or test-time-augmentation aggregation, which combines raw per-view representations rather than differences between them: aggregating perturbed views without differencing is exactly the Φ(xi+δi)\Phi(x_i+\delta_i) row above, and it does not recover the gain.

(4) Latency and peak memory.

Measured on a single RTX 4090 with bf16 at batch size 1 and 224×224224\times224 input: one encoder pass takes 27.8,ms27.8\\,\mathrm{ms} and peaks at 12.60,GiB12.60\\,\mathrm{GiB}, of which 12.52,GiB12.52\\,\mathrm{GiB} is the frozen backbone. One-perturbation CPR takes 56.1,ms56.1\\,\mathrm{ms} and R∗R^{\ast} takes 112.3,ms112.3\\,\mathrm{ms}, with no increase in peak memory, since responses are differenced per branch. Memory scales with batch size and resolution; at 518×518518\times518 with batch size 16, the peak reaches 20.1,GiB20.1\\,\mathrm{GiB} (22.6,GiB22.6\\,\mathrm{GiB} per nvidia-smi). A single FFT branch needs one forward pass, the same as DINOv3-Linear, and stacking multiple linear heads did not improve the DINOv3-Linear baseline in our runs.

Reviewer_sy6QOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Carrier-Perturbation Responses (CPR) to detect AI generated images by measuring feature changes between original and perturbed images. The authors further introduce an ANOVA-inspired perturbation selection criterion to select suitable perturbations from three perspectives. Experimental results show that the proposed method achieves good detection performance on several cross domain datasets.

Critical strengths

The paper proposes a variance-based perturbation selection method to select effective perturbations from different categories and combines multiple perturbation responses for detection. This method shows some novelty and achieves good performance in cross domain AI generated image detection.

Critical weaknesses
  1. The paper presents CPR as a new basis for image authenticity assessment. However, existing studies, such as RIGID and RA-Det, have also explored detecting AI-generated images through feature changes induced by image perturbations. The paper does not sufficiently discuss these closely related works or clearly explain the key differences between CPR and existing methods. Therefore, I think the novelty of the proposed method is not clearly presented.
  2. The authors argue that a stronger carrier-perturbation interaction helps select more effective perturbations, but this assumption is not sufficiently justified. For example, images from different scenes may respond very differently to the same perturbation, resulting in a high interaction effect. However, such differences may mainly come from image content rather than the differences between real and generated images.
Questions for authors

Why does a higher carrier-perturbation interaction indicate better separability between real and generated images? For example, if landscape images and portrait images respond very differently to the same perturbation, the interaction effect may mainly reflect scene differences rather than image authenticity. How do the authors distinguish between these two cases? Are there any experiments that support this design choice?

4 comments
AuthorsOpenReview ↗

Key Differences from RIGID and RA-Det

We agree that these are the closest works and that they were not adequately discussed. We consider the essential differences to concern what the perturbation response is and what it is used for, rather than the choice of backbone or benchmark.

RIGID adds isotropic Gaussian noise and compares cos⁡(f(x),f(x+λδ))\cos(f(x), f(x+\lambda\delta)). The response is collapsed to a single scalar and compared against a quantile threshold calibrated on real images. Two differences are essential rather than incremental.

  1. Scalar versus vector response. RIGID retains only the magnitude of the response. CPR retains the full response vector, and our criterion (§ 3.2) selects precisely on the structure of that vector across a rule ensemble. No scalar can represent this information.

  2. Unstructured versus structured probes. Isotropic Gaussian noise has no structure for the carrier to modulate. In our decomposition such a probe places almost all response variance in the carrier and rule terms, and the interaction term stays near zero. In our own measurements this is the regime the criterion rejects (ρF,1\rho_{F,1} below, ηcar=0.728\eta_{\mathrm{car}} = 0.728, ηint=0.254\eta_{\mathrm{int}} = 0.254). Structured probes such as frequency-band filtering, patch-wise noise and attention-guided masking are what make the interaction channel informative in the first place. This is why a selection criterion over perturbation categories is needed at all. RIGID has no analogue of one, because it has a single perturbation family with a single intensity parameter.

RA-Det uses a learned perturbation generator trained jointly with a multi-branch detector under Lbce+ReLU((sfake−sreal)+γ)\mathcal{L}_{\mathrm{bce}} + \mathrm{ReLU}\big((s_{\mathrm{fake}} - s_{\mathrm{real}}) + \gamma\big). The essential difference is in how the probe itself is supervised.

  1. Supervision of the probe. RA-Det's probe is optimised against authenticity labels so that generated images drift more than real ones. The asymmetry it measures is therefore partly induced by the objective, and the probe inherits the ProGAN training domain. CPR's perturbations are analytically given and never optimised against labels. CPR assumes neither a direction (real images being the more stable class) nor a generator. The two works make different claims. RA-Det shows that a label-supervised probe can manufacture a usable asymmetry. CPR asks which analytically defined probe makes the response structure depend on the carrier rather than on the rule.

On the publication timeline, RIGID appeared in TMLR in August 2026. The ICML conference at which RA-Det was presented was held in July, and its proceedings volume was formally published on 29 September, slightly after the 25 September submission deadline. We will add both references and clarify the distinctions above. In light of this timeline, we would respectfully ask the reviewer to regard the current insufficient discussion as a matter that can be remedied in revision.

AuthorsOpenReview ↗

Scene Effects and Probe Selection

Rationale.

We thank the reviewer for raising this concern. The key distinction is whether scene-related response differences persist across perturbation rules or vary with the rule. Different scenes may respond differently to a given perturbation, but this alone does not mean that the difference dominates the interaction. Under our decomposition, a scene-related difference that remains approximately constant across rules contributes to the carrier effect, whereas the component that varies across rules may contribute to the carrier–rule interaction. We therefore do not assume that the interaction is free of scene effects.

Our motivation for prioritizing interaction is to give greater relative emphasis to rule-dependent response differences beyond persistent carrier offsets. When scene variation is expressed mainly as such offsets, this preference is intended to make probe selection less dependent on persistent scene differences. This is a selection rationale, not a claim that interaction directly measures authenticity or is scene-invariant. As specified in Appendix A, variance statistics are computed separately for real and generated images and then equally averaged. Eq. (13) jointly considers interaction, carrier and rule effects, and response magnitude; the carrier effect receives a positive weight. Classification uses the full CPR vector. We evaluate the practical value of this joint criterion below.

Additional evidence.

The table below reports two complementary evaluations. First, ForenSynths/ProGAN contains 20 semantic categories, each with 200 real and 200 generated images; AUROC is computed separately within each category. Second, we vary the real-image source among COCO, OpenImages, and ImageNet while keeping the generated set fixed. Both evaluations use source-estimated real/fake centroids for scoring, with no target-data fitting. For each family, we compare the CPR-selected probe with the maximum-carrier-effect candidate identified in the original Table 2.

AUROC (%). ProGAN reports mean ±\pm SD across 20 categories; other columns correspond to different real-image sources. Within each pair, ∗\ast marks the CPR-selected probe; the other is the maximum-carrier-effect control.

ProbeProGANCOCOOpenImagesImageNet
ρF,1\rho_{F,1}74.33±5.5674.33 \pm 5.5685.882.683.3
ρF,3∗\rho_{F,3}^{\ast}99.25±0.9599.25 \pm 0.9595.593.494.2
ρS,2\rho_{S,2}38.53±13.2138.53 \pm 13.2186.983.681.6
ρS,4∗\rho_{S,4}^{\ast}98.02±1.8698.02 \pm 1.8693.491.895.0
ρA,1\rho_{A,1}97.17±1.9097.17 \pm 1.9086.085.786.3
ρA,4∗\rho_{A,4}^{\ast}97.98±1.2997.98 \pm 1.2990.988.890.3

The within-category evaluation directly tests authenticity discrimination while holding semantic category fixed. The selected frequency and structure probes outperform their paired controls in all 20 ProGAN categories, and the attention probe does so in 17 of 20 categories. Thus, the selected responses distinguish real from generated images within the same semantic category, and their advantage cannot be attributed solely to differences in semantic-category composition. All three selected probes also outperform their controls for each real-image source, supporting their discriminative value under changes in the real-image domain.

These diagnostics complement Table 5, which evaluates all 11 candidates on Chameleon, WildRF, SocialRF, and SynthWildX. The selected probes achieve average accuracies of 94.6%, 94.1%, and 93.8%, occupying the top three positions among all candidates, compared with 78.9–91.7% for the remaining eight. Together, the category-controlled evaluation, the real-image source-change evaluation, and the cross-domain comparison provide complementary evidence that the interaction-prioritized joint criterion selects probes with transferable authenticity-discriminative information. These findings support the selection rationale beyond an explanation based solely on semantic-category differences.

Thanks for the detailed response. The authors have addressed my concerns and provided the evidence I asked for. I will raise my score to 3.

Thank you for carefully considering our responses and the additional evidence, and for updating your assessment of our work. We are very glad that our clarifications and experiments have addressed your concerns. We sincerely appreciate the time and thought you have devoted to reviewing our paper and engaging with us throughout the discussion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We have revised the manuscript to incorporate the corrections, additional experiments, and clarifications presented during the discussion. We have highlighted the updates in red. The main updates include:

  1. Corrected the reported results and dataset statistics in Tables 4 and 8.

  2. Added references to SigLIP2, DINOv2, RIGID, and RA-Det.

  3. Clarified the positioning of CPR and discussed its contributions and scope relative to existing perturbation-based detectors.

  4. Extended the evaluation to multiple pretrained backbones beyond DINOv3 in Section 4.5 and Table 7.

  5. Added inference-cost measurements, including the forward-pass budget, inference latency, and peak memory, in Appendix H.

  6. Added category-controlled and real-image source-change experiments in Appendix E.

  7. Added cross-dataset validation of the unchanged selection criterion on OpenImages in Appendix F.

We extend our deepest gratitude to all the reviewers for their time and valuable feedback.

AuthorsOpenReview ↗

Author Summary for the AC and Reviewers

Dear Reviewers and Area Chair,

Thank you for the time and care devoted to evaluating our paper. We summarize the main results below.

Main Contributions

Our paper, “Carrier-Perturbation Responses: An Alternative Basis for Generalizable AI-Generated Image Detection,” explores how images respond to controlled perturbations as a basis for authenticity assessment. Our main contributions are:

  1. Perturbation-Response Representation. We introduce CPR, which uses perturbation-induced feature changes to capture authenticity cues beyond an image’s original features.

  2. ANOVA-Inspired Perturbation Selection. We develop a criterion that jointly considers variance components and response magnitude to select effective perturbations.

  3. Cross-Domain Detection with Lightweight Training. We demonstrate improved cross-domain detection using a frozen encoder and linear classifiers, with additional experiments supporting the usefulness of CPR beyond DINOv3.

These contributions received positive recognition from the reviewers:

  • Value of the response representation: “Characterizing how an image's representation changes under a controlled perturbation, rather than relying exclusively on its unperturbed features, is an intuitive and potentially useful direction.” (Reviewer gceq)

  • Novelty of perturbation selection and combination: “The paper proposes a variance-based perturbation selection method to select effective perturbations from different categories and combines multiple perturbation responses for detection. This method shows some novelty and achieves good performance in cross domain AI generated image detection.” (Reviewer sy6Q)

  • Simplicity and supporting ablation: “Simple, practical idea with a clear ablation.” (Reviewer NGsF)

  • Cross-domain effectiveness: “CPR improves over the same-backbone DINOv3-Linear baseline on all four datasets, with the clearest gains in real-image accuracy.” (Reviewer NGsF)

Summary of the Reviewer Discussions

We thank the reviewers for their constructive feedback. Below, we summarize the specific issues addressed during the discussion.

  1. Reviewer sy6Q questioned whether the carrier–perturbation interaction might mainly reflect scene differences. We clarified this issue and provided category-controlled and real-image source-change evidence supporting the discriminative value of the selected responses beyond semantic-category differences. The reviewer acknowledged that our response addressed the concerns and raised the score to 3.

  2. Reviewer gceq raised concerns about the evaluation protocol, perturbation-selection validation, and reproducibility. We clarified the evaluation and aggregation procedures, provided additional validation of the fixed selection criterion, and supplied the requested implementation and computational details. The reviewer thanked us for the clarifications and additional evidence.

  3. Reviewer NGsF questioned whether the performance improvements mainly originated from the DINOv3 backbone rather than from CPR itself. We clarified the respective roles of the pretrained backbone and the CPR representation, and provided additional evidence from PE-CLIP, SigLIP2, CLIP, and DINOv2-giant. These results support the usefulness of CPR beyond DINOv3 while also revealing how its benefits vary across backbones and evaluation domains. The reviewer confirmed that our response addressed the concern and raised the score to 3.

These exchanges strengthened the experimental support, improved evaluation transparency, and helped us define the contribution and scope of CPR more precisely.

We are grateful once more for the engaging and constructive review process, which has undeniably enriched our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.