acceptodds
Under review as a conference paper at ICLR 2027

No Shortcuts to Truth: Exposing and Eliminating Shortcuts in AI-Generated Image Detection via Cross-Generator Consistency

Abstract

Current synthetic image detectors frequently saturate on training generators but collapse on unseen ones. We trace this generalization failure to shortcut learning, a phenomenon where models capture spurious dataset correlations rather than genuine forgery evidence. We identify a fundamental structural asymmetry between these two feature types. Authenticity evidence is strictly confined to one side of the real and fake boundary, whereas shortcuts can manifest in both classes. Consequently, a shortcut-dependent detector yields fluctuating verdicts on the exact same real image when the paired synthetic generator changes. In contrast, an evidence-based model maintains prediction consistency. Building upon this insight, we propose Real Anchored Invariance, a novel data efficient multiple source framework that eliminates shortcuts without requiring prior knowledge of specific artifacts. Our method penalizes environment specific residuals along high variance directions of shared real features, dynamically upweights authentic instances that expose prediction disagreements, and simulates generator-level distribution shifts via episodic training. Extensive evaluations demonstrate that our approach achieves exceptional robustness and transferability across diverse unseen benchmarks and strictly held-out cutting-edge generators, even under extreme data scarcity.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Image Detection, DeepFake

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_qv6aOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RAIN, a framework for improving the generalization of AI-generated image detection by mitigating shortcut learning. It introduces three components: Real-Anchored Alignment, Disagreement-Guided Reweighting, and Leave-Environments-Out optimization. Experiments across multiple benchmarks demonstrate strong generalization to unseen generators and robustness under limited training data.

Critical strengths
  1. Interesting perspective on shortcut learning. The paper provides a well-motivated analysis of acquisition, content, and style shortcuts, supported by controlled experiments demonstrating their impact on generalization. Using cross-generator consistency on shared real images as a training signal is an interesting approach that avoids explicitly specifying individual shortcut attributes.

  2. Strong empirical performance and comprehensive evaluation. RAIN achieves substantial improvements over competitive few-shot baselines across in-domain, out-of-domain, and strictly held-out generator evaluations. The component ablations and corruption robustness experiments provide additional evidence of the framework's effectiveness, particularly under limited training data.

Critical weaknesses
  1. Insufficient justification for the connection between prediction consistency and shortcut elimination. The central assumption is that disagreement among generator-specific classifiers on real images indicates shortcut reliance. However, such disagreement may also arise from legitimate differences in learned decision boundaries, limited training samples, or optimization variability. Conversely, different classifiers may consistently exploit the same spurious features without exhibiting disagreement. Therefore, enforcing cross-generator consistency does not necessarily guarantee shortcut removal. The current analysis demonstrates improved prediction agreement but does not sufficiently establish that the learned representations are genuinely less dependent on spurious correlations.

  2. Limited isolation of the proposed mechanism from other contributing factors. Although the component ablations demonstrate performance improvements, the evaluation does not adequately separate the benefits of shortcut suppression from those of the pretrained DINOv3 backbone, multi-source training, and additional regularization. In particular, the controlled shortcut experiments primarily establish that conventional detectors are vulnerable to injected biases, rather than directly measuring whether RAIN eliminates these biases. More targeted interventions and backbone-matched comparisons would strengthen the causal interpretation of the results.

Questions for authors
  1. How can the authors establish that cross-generator prediction consistency specifically suppresses shortcuts? Could the authors evaluate RAIN under the controlled acquisition, content, and style shortcut interventions introduced in Section 3, comparing against standard ERM and other invariance-based objectives?

  2. Can the authors clarify whether prediction consistency is sufficient to identify shortcut reliance? In particular, how does RAIN handle spurious features shared across multiple generator environments, where different classification heads may agree despite relying on the same shortcut?

  3. How much of the improvement comes from the proposed invariance mechanism rather than the pretrained representation? A comparison with backbone-matched ERM and alternative regularization methods under identical training budgets would help clarify the contribution of RAIN.

5 comments

Dear Reviewer #qv6a, we sincerely thank you for your precious time and constructive comments. We are deeply encouraged by your positive recognition of our interesting perspective on shortcut learning, strong empirical performance, and comprehensive evaluation. We sincerely hope the following clarifications and new experiments can address your concerns.


W1.1: Cross-head disagreement on real images may also come from legitimate boundary differences, limited samples, or optimization variability, rather than shortcuts.

A1.1: Thank you for this insightful comment! We fully agree that these factors also cause disagreement. However, shortcut-induced disagreement is much larger, and RAIN is designed to suppress only this part.

  • RAIN does not penalize all disagreement. For legitimate boundary differences, each head naturally learns the fake-specific artifacts of its own generator, such as GAN upsampling patterns or VAE decoding traces. These artifacts are absent from real images, so real features barely vary along these directions, which correspond to low-variance directions of ΣR\Sigma_R. Since RA penalizes the head residual δe=we−wˉ\delta_e = w_e - \bar{w} through δe⊤ΣRδe\delta_e^\top \Sigma_R \delta_e, these differences are hardly penalized, and each head can still use them to detect its own generator. For disagreement from limited samples and optimization variability, DGR only up-weights real images with large disagreement, so noise-level disagreement receives little extra supervision.
  • Following your suggestion, we quantify the two sources of disagreement. We construct 10 real–fake paired environments sharing the same real images and inject an ACS shortcut into the fake images of 3 of them. We train a multi-source ERM and report the disagreement rate, which is the fraction of unseen real images on which the heads' predictions are inconsistent.
Unseen realDisagreement rate
Clean0.278
+ Acquisition0.783
+ Content0.926
+ Style0.583

From the results, we find that disagreement from limited samples and optimization noise does exist, but shortcuts cause substantially larger disagreement across all three shortcut families. We will add these results and discussion in our revision.

W1.2: Heads may consistently rely on the same spurious feature without any disagreement.

A1.2: Thank you for this observation!

  • We agree that an attribute shared by all environments with the same intensity cannot be separated from authenticity evidence without external prior knowledge. This limits any environment-based mitigation strategies, since both are equally predictive in every environment.

  • However, in practice, a shared shortcut rarely has the same intensity across generators, since their pipelines differ. The heads depend on it to different degrees, and the disagreement signal remains.

  • Following your suggestion, we inject the same shortcut into all training environments under two settings with the same total number of shortcut-carrying samples: uniform (identical intensity across environments) and non-uniform (varying intensity). We report the shortcut reversed AUC.

Shared shortcutUniform (ERM)Uniform (RAIN)Non-uniform (ERM)Non-uniform (RAIN)
Acquisition0.0860.4360.0850.945
Content0.0160.7110.1300.964
Style0.0030.3040.0020.851

ERM fully relies on the shared shortcut in both settings. When the shared shortcut differs in intensity across environments, RAIN exploits this difference and largely recovers. When the intensity is perfectly identical, the cross-head signal vanishes and RAIN's gain is limited, although it still outperforms ERM. This is a common boundary of environment-based invariance learning, and we will discuss it as a limitation in our revision.

W1.3: Improved agreement does not show that the learned representations rely less on spurious correlations.

A1.3: Thank you for this important point! We measure reliance directly with the Hard AUC, which is independent of cross-head agreement. Under the same setup as A1.1, we compare ERM and RAIN.

Unseen realHard AUC ↑ (ERM)Hard AUC ↑ (RAIN)
Without shortcut attribute0.8380.982
+ Acquisition0.3520.979
+ Content0.3360.878
+ Style0.3770.926

ERM falls below chance once the shortcut is reversed, indicating that it largely decides based on the shortcut. RAIN stays well above chance across all three shortcut families, showing that its decisions rely substantially less on spurious correlations, rather than merely becoming more consistent. We provide per-image disagreement maps in the anonymous repository and will add these results in our revision.


W2.1: The evaluation does not separate the benefit of shortcut suppression from those of the DINOv3 backbone, multi-source training, and additional regularization.

A2.1: Thank you for this constructive suggestion! We isolate each factor under the same data and training budget as follows.

  • Pretrained backbone. We replace the encoder and compare RAIN with backbone-matched multi-source ERM. The AUC results are shown below.
EncoderERMRAIN (ours)Δ
DINOv3 ViT-L/160.9140.999+0.085
DINOv2 ViT-L/140.8180.981+0.164
DINOv2 ViT-B/140.8030.988+0.185
CLIP ViT-L/140.9360.972+0.036
ResNet-500.7740.840+0.066

RAIN consistently improves over the backbone-matched ERM on all five encoders, including the convolutional ResNet-50, so its gain is not tied to DINOv3 representations.

  • Multi-source training. We respectfully clarify that Table 4 of our paper already isolates this factor under the same DINOv3 + LoRA backbone.
SettingAP ↑FNR ↓
Baseline 1: ERM on ProGAN data71.345.4
Baseline 2: ERM on the same 5-shot multi-source data as RAIN83.758.6
Baseline 2 + RA99.620.4

Multi-source data alone improves AP but also increases FNR. With identical backbone, data, and training budget, adding RA alone brings the major improvement, so the gain beyond Baseline 2 comes from our mechanism rather than from multi-source training.

  • Additional regularization. Starting from the multi-source ERM with DINOv3 ViT-L/16, we add generic regularizers and a head-consistency penalty that applies an L2 penalty to the δe\delta_e of the heads. The AUC results are shown below.
MethodAUC ↑
ERM0.914
+ Dropout (0.5)0.916
+ Label smoothing0.911
+ head L20.904
RAIN (ours)0.999

Generic regularizers bring negligible changes and the head L2 penalty even degrades performance. Together, these results confirm that RAIN's improvement comes from its invariance mechanism rather than from the pretrained backbone, multi-source training, or extra regularization. We will add these results and discussion in our revision.

W2.2: The controlled shortcut experiments only show that conventional detectors are vulnerable, without directly measuring whether RAIN eliminates these biases.

A2.2: Thank you for this comment! Please kindly refer to A1.3 and A3, where we directly evaluate RAIN under the controlled ACS shortcut interventions against ERM and other invariance-based objectives.

Q1: Can RAIN be evaluated under the controlled ACS shortcut interventions of Sec. 3, against ERM and other invariance-based objectives?

A3: Thank you for this constructive question! Following your suggestion, we further compare RAIN with IRMv1, V-REx, and GroupDRO. All methods use the same encoder, multi-head, 5-shot data from AIGIBench environments with shared real images, and training budget, with the penalties of the baselines applied to the per-environment head risks. Following the operators in Table 6, we inject the same acquisition shortcut (JPEG) into the fake images of k environments (k = 1 or 12), and report the AUC under the shortcut-removal (Equal) and shortcut-reversal (Hard) protocols of Sec. 3.

kMethodEqual ↑Hard ↑
1ERM0.8530.296
1IRMv10.8540.343
1V-REx0.8520.305
1GroupDRO0.8530.297
1RAIN (ours)0.9790.891
12ERM0.8180.056
12IRMv10.8060.107
12V-REx0.8130.057
12GroupDRO0.8180.056
12RAIN (ours)0.9400.894

Existing invariance objectives behave almost identically to ERM and fall far below chance under reversal, since per-environment training risks quickly approach zero in the few-shot setting, leaving little signal for risk-based penalties. RAIN instead anchors on the predictions of shared real images and remains robust, even when the shortcut is shared by about half of the environments. We will add these results in our revision.


Q2: Is prediction consistency sufficient to identify shortcut reliance, especially when spurious features are shared across environments?

A4: Thank you for this question! Please kindly refer to A1.2.


Q3: How much of the improvement comes from the invariance mechanism rather than the pretrained representation?

A5: Thank you for this question! Please kindly refer to A2.1, where we compare RAIN with backbone-matched ERM and alternative regularization under identical training budgets.

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

This paper studies shortcut learning in AI-generated image detection and proposes Real-Anchored Invariance (RAIN) to improve generalization to unseen generators under a few-shot multi-source setting. The framework uses a shared DINOv3 encoder adapted with LoRA and a linear classification head for each generator environment. Real-Anchored Alignment (RA) penalizes environment-head residuals along directions with high real-feature variance. Disagreement-Guided Reweighting (DGR) increases supervision on real samples that produce substantial disagreement across standardized head logits. Leave-Environments-Out (LEO) differentiates through a temporary ridge classifier fitted on a subset of generators and evaluated on the remaining generators. At inference, the model retains the encoder and a test head using the mean environment-head weight with an independently learned bias. Evaluations cover AIGIBench, four external benchmarks, fourteen held-out generators, and image corruptions.

Critical strengths
  1. Shortcut reliance is an important obstacle to reliable image detection. The acquisition, content, and style taxonomy provides an accessible organization of this problem. The controlled experiments involving preserved, removed, and reversed shortcuts help demonstrate how strong in-distribution performance can conceal unreliable decision rules.

  2. RA constrains head disagreement through the real-feature covariance rather than requiring all head parameters to coincide. This allows environment-specific directions with low variation on real images to remain available for discrimination. DGR complements this distribution-level constraint with instance-level supervision, while LEO explicitly encourages transfer between generator environments.

  3. The paper evaluates multiple sample budgets, heterogeneous forgery families, external datasets, unseen generators, and several corruption types. The reported performance under very small support sets is promising. The component ablation also provides initial evidence that the proposed training components contribute beyond naive multi-source training.

Critical weaknesses
  1. Missing comparisons with directly relevant data alignment methods leave the advantages of RAIN insufficiently established. The paper motivates RAIN partly through the limitations of existing data alignment approaches, yet does not compare against directly relevant methods such as DDA, AlignedForensics, DRCT, or B-Free. In particular, DDA already investigates frequency biases introduced by reconstruction and mitigates them through pixel- and frequency-domain alignment; its effects are not limited to individually addressing explicitly named shortcuts. Consequently, the current experiments do not establish RAIN’s additional benefits over these approaches or substantiate stronger mitigation of shortcuts that are not explicitly targeted. Direct comparisons with matched backbones, data sizes, and training budgets are needed, together with an analysis of whether RAIN still provides gains on already aligned data.

  2. Cross-environment consistency does not guarantee shortcut elimination, and the proposed mechanism lacks direct validation. If a label-correlated shortcut is shared by all training generators, the classification heads can consistently rely on it. RA primarily constrains the residuals of environment heads relative to their mean weight, so shared shortcuts may persist in the average classification direction. DGR also provides little corrective signal when the heads agree, while LEO may continue to exploit biases shared across training environments. The controlled ACS experiments demonstrate that conventional detectors learn shortcuts, but do not directly evaluate RAIN under the same shortcut-removal and reversal protocols. Increased prediction correlation across heads does not, by itself, establish that the model has shifted toward authenticity evidence. The current evidence therefore supports reducing some environment-dependent shortcut reliance more strongly than the broader claim of shortcut elimination.

  3. The stability of the extreme few-shot results and the estimation of key statistics remain unclear. The paper emphasizes strong performance in the one-shot setting, but does not report performance variability across independently sampled support sets. Further clarification is also needed regarding the number of unique real images used at each shot budget and how the real-feature covariance and head-logit standard deviations are estimated in the one-shot setting. These statistics directly contribute to RA and DGR, while extreme few-shot performance may be sensitive to support-sample selection. The current reporting is therefore insufficient to assess whether the performance advantages are stable and whether the real-anchored constraints have a reliable statistical basis in this setting.

  4. The risk of LEO overfitting to the finite set of training environments is insufficiently analyzed. Within LEO, held-out generators are excluded only from fitting the temporary classifier in the current episode; their samples still update the shared encoder through the outer loss. Consequently, the objective may learn cues covering the known generators without necessarily producing evidence that transfers to entirely new generators. The results on external datasets and unseen generators provide positive evidence, but the generator-validation protocol used for model selection and the effect of training duration remain unclear. Validation on completely separate generators, evaluations holding out entire generator families, or training-duration analyses would help characterize LEO’s transfer benefits and limitations.

Questions for authors
  1. Could you provide direct comparisons with data alignment methods such as DDA, AlignedForensics, and DRCT to clarify RAIN’s advantages over these approaches?

  2. How does RAIN handle shortcuts shared by all training generators? Could you provide controlled experiments that directly validate its shortcut mitigation mechanism? Please clarify how key statistics are estimated in the extreme few-shot setting and describe the generator-based validation protocol used for hyperparameter and checkpoint selection.

  3. The problem is important and RAIN is promising, but the evidence for shortcut elimination remains insufficient. Unclear evaluation protocols and reporting inconsistencies further limit confidence. Stronger controlled experiments and corrected reporting could change my assessment.

  4. Minor formatting issue: The page headers still refer to ICLR 2026. Please update them to ICLR 2027.

6 comments

Dear Reviewer #bzxH, we sincerely thank you for your precious time and valuable comments. We are deeply encouraged by your positive recognition of the importance of our research problem and the promising performance. We sincerely hope the following clarifications and new experiments can address your concerns.


W1: Direct Comparisons with Data Alignment Methods

A1: Thank you very much for this constructive suggestion! We fully agree that direct comparisons are needed to establish RAIN's benefits over data alignment.

  • We first clarify the relation between RAIN and data alignment. AlignedForensics and B-Free reconstruct real images to obtain content-aligned fake images. DDA further finds that pixel-domain and frequency-domain discrepancies remain after reconstruction, and performs dual alignment on both. DRCT uses reconstructed real images, original fake images, and reconstructed fake images as hard negatives for contrastive learning. The data alignment pipelines are essentially still generative processes, which may introduce new latent shortcuts. Moreover, they require reconstructing a large number of real samples, and the alignment modifications are fixed operations. RAIN provides a new structural perspective that requires neither fixed alignment operations nor diffusion-model reconstruction. Since shortcut attributes form an open set, perfect data alignment cannot be achieved. We therefore construct paired environments in the multi-source few-shot setting to mitigate the reliance on shortcuts.

  • Following your suggestion, we add direct comparisons under the same data size and training cost. Specifically, we obtain fake images with the official reconstruction models of AlignedForensics, DDA, and DRCT, and merge them with the same fake samples as RAIN to form the training set for retraining. The training cost is unified to 160, 800, and 1,600 iterations for 1-shot, 5-shot, and 10-shot, respectively. The AUC results are shown below.

MethodShotAIGIBenchGenImageChameleonAIGCDBCO-SPY
DDA10.4540.4380.5740.4480.488
DRCT10.7270.6720.6390.6020.598
AlignedForensics10.7470.6960.7150.6290.642
RAIN (ours)10.9830.9890.9170.9910.986
DDA50.9350.9300.7090.9410.787
DRCT50.8930.8440.7120.7750.777
AlignedForensics50.8770.8820.7630.7820.813
RAIN (ours)50.9990.9900.9400.9880.998
DDA100.9440.9610.6510.9740.822
DRCT100.9230.9170.7900.7920.854
AlignedForensics100.9020.9130.7810.8000.879
RAIN (ours)100.9990.9970.9610.9930.998

The data alignment methods perform poorly in this setting, as they are designed for large-scale reconstruction. Under the same setting, RAIN also surpasses the dedicated multi-source few-shot detectors, such as FSD, FTNet, and Fleet, as shown in Tables 1–3.

  • To avoid understating these methods, we further compare with their official full-scale checkpoints. RAIN still achieves better performance with a much smaller training cost. This indicates that our advantage stems from the training signal rather than data volume.
MethodWeightsTraining dataAIGIBenchGenImageChameleonAIGCDBCO-SPYAIGenImages2026
DDAOfficial236k0.9230.9730.9190.9470.9840.938
DRCTOfficial359k0.8500.9800.7460.8170.9740.776
AlignedForensicsOfficial472k0.7750.8950.8060.7460.9880.888
RAIN (ours)10-shot0.52k0.9980.9970.9610.9950.9980.978

We will include both comparisons and the complementarity analysis in the revision. RAIN weights and retraining scripts are provided in the anonymous repository. Thank you again for this valuable suggestion!

W2.1: If a label-correlated shortcut is shared by all training generators, the classification heads can consistently rely on it. How does RAIN handle shortcuts shared by all training generators?

A2.1: Thank you for this insightful point.

  • We first clarify the scope of RAIN that the consistent signal of RAIN comes from differences between generator environments. When a label-correlated attribute appears in only some environments or with different intensity, the heads of these environments rely on this attribute to different disgrees. Heads therefore disagree on real images that contain the attribute, and fitting on these environments fails to transfer to the others. RA and DGR use this cross-head disagreement, and LEO uses this transfer failure, as signals to suppress the reliance on the attribute.
  • We agree that an attribute shared by all environments with the same intensity cannot be separated from authenticity evidence without external prior knowledge. This limits any environment-based mitigation strategy, since the shortcut and authenticity evidence are equally predictive in every environment.
  • In practice, a shared attribute rarely has the same intensity across generators, since their pipelines differ. To verify RAIN exploits this intensity variation against shortcuts, we construct two training settings with the same total number of shortcut-carrying samples. In the uniform setting, all environments carry the shortcut with the same intensity, and in the non-uniform setting, the intensity differs across environments. The shortcut reversed AUC is reported below.
ShortcutUniform ERMUniform RAINNon-uniform ERMNon-uniform RAIN
Acquisition0.0860.4360.0850.945
Content0.0160.7110.1300.964
Style0.0030.3040.0020.851

When the intensity varies, RAIN restores a high reversed AUC, while ERM still follows the shortcut. Therefore, RAIN handles shortcuts whose intensity varies across generators, which is the common case in real-world scenarios. Uniformly shared shortcuts call for more diverse environments or data-level alignment, both complementary to RAIN. We will add this experiment and discussion to the revision.

W2.2: The controlled ACS experiments do not directly evaluate RAIN under the same removal/reversal protocols, and increased prediction correlation alone does not establish a shift toward authenticity evidence. Could you provide controlled experiments that directly validate its shortcut mitigation mechanism?

A2.2:

  • Following your suggestion, we evaluate RAIN under the identical Equal / Hard protocols of Sec. 3.2, using the round-1 operators of Table 6 and the same conventional detector as reference. The reported metric is AUC.
ShortcutEqual (removed)Equal (removed)Hard (reversed)Hard (reversed)
ERMRAINERMRAIN
Acquisition0.5170.9560.3070.904
Content0.6290.9240.5510.964
Style0.5770.9610.3310.937

RAIN discriminates real from fake once the shortcut is removed. Its Equal AUC stays high for all three families, whereas the baseline drops close to chance. The Hard AUC of RAIN indicates that its decisions no longer depend on these shortcuts. We will add these results to Sec. 3.2 in our revision.

W2.3: The current evidence therefore supports reducing some environment-dependent shortcut reliance more strongly than the broader claim of shortcut elimination.

A2.3: We revise it to "reduce / mitigate" throughout, and state the precise claim "Eliminating all shortcuts is this field's broader goal and RAIN contributes a shortcut-agnostic, training objective-level signal that reduces shortcut reliance, complementing data-level alignment". Furthermore, appendix F.2 documents four alternative mitigation strategies we explored to situate RAIN in this design space.

W3.1: Performance variability across independently sampled support sets is not reported.

A3.1: Thank you for this careful comment. Following your suggestion, we retrain RAIN at 1 / 5 / 10 shots with five independently sampled support sets and evaluate the last-epoch weights, reporting AP as mean ± std. The results confirm that the real-anchored constraints retain a reliable statistical basis. Per-seed values, settings and checkpoints are in the anonymous repository referenced in the reproducibility statement in lines 499 to 505, and we are happy to assist with any reproduction issue.

Benchmark1-shot5-shot10-shot
AIGIBench97.38 ± 0.6199.20 ± 0.1899.38 ± 0.16
GenImage96.98 ± 2.3499.38 ± 0.6799.56 ± 0.31
AIGCDB98.58 ± 1.0699.64 ± 0.49100.00 ± 0.00
Chameleon89.96 ± 5.0995.56 ± 2.5996.62 ± 2.36
CO-SPY93.07 ± 1.9597.87 ± 1.3198.40 ± 1.49

W3.2: The number of unique real images per budget, and how the real-feature covariance and head-logit standard deviations are estimated in the 1-shot setting.

A3.2: We clarify the data and estimation protocol, and they ensure that neither statistic is estimated from a single image.

  • We use 26 environments, the 25 generators in Table 1 plus SD1.4. Following Sec. 3.1, each environment contributes n fake images and n real images, so the shared real pool holds 26n unique real images. Thus even the 1-shot setting trains on 26 distinct real images, not one. The training and test splits are strictly disjoint under fixed-seed sampling.
  • For RA, μR\mu_R is a running mean across steps and ΣR\Sigma_R is recomputed at every step from the centered 512-d features of the current real batch under fresh random cropping and JPEG augmentation. For DGR, d(xR)=Vare[s^e(xR)]d(x_R)=\mathrm{Var}_e[\hat{s}_e(x_R)] is the variance across the 26 environment heads on a single real image and is therefore independent of the shot budget. The standardization in Eq. 32 uses an EMA running mean across steps and the current real-batch standard deviation, both accumulated over training.

W4.1: Held-out generators still update the encoder through the outer loss, so LEO may learn cues covering the known generators rather than evidence that transfers to entirely new ones.

A4.1:

  • The bilevel design prevents memorizing known generators. In each episode the temporary head hSh_S is fitted from scratch on SS only and is discarded after the outer update, so the encoder cannot store a per-generator decision rule. The outer loss evaluates this SS-fitted head on QQ, hence the only way to reduce it is to make directions discriminative on SS equally valid on the fakes of QQ. A cue that holds for only part of the generators fails in every episode where those generators fall into QQ, and is penalized as random partitions accumulate. Updating the encoder on QQ is therefore the intended transfer signal, not a shortcut around the holdout.

  • This is borne out empirically. Table 3 reports consistently high accuracy across all 14 strictly held-out generators, none of which appears in any support set. Fig. 5(c) provides the family-holdout evaluation the reviewer suggests, GAN-6 and Diff-7 train on a single family and are tested on all five, so four families are entirely unseen. Adding LEO alone to the multi-source baseline also substantially reduces the false-negative rate (Table 4), confirming its transfer contribution.

W4.2: The generator-validation protocol used for model selection and the effect of training duration remain unclear.

A4.2: Thank you for raising these points! We clarify them item by item below. This response also covers the related parts of Q2 and Q3.

  • Checkpoint selection and training duration: Training uses a fixed budget of 160, 800, and 1,600 steps at 1, 5, and 10 shots. We save a checkpoint every 40 steps, evaluate each one on the in-domain and out-of-domain benchmarks, and report the best one in Sec. 5. We agree that this selection relies on the evaluation benchmarks, since no separate validation generators are held out.
  • Strictly held-out generators. The 14 generators in Table 3 are never used for checkpoint selection.
  • Hyperparameters. λRAλ_{RA} and λLEOλ_{LEO} are set by the ablation in Fig. 5(a)(b). We therefore fix λRAλ_{RA}= λLEOλ_{LEO}= 1 for all budgets and benchmarks. All other hyperparameters are likewise shared across experiments and will be listed in appendix D.3.
  • Corrected reporting. We corrected the swapped panel titles of Fig. 6, the caption of Table 6 and the page header. The main conclusions are unchanged.
Reviewer_GU5WOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3(+1)
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies shortcut learning in AIGI detection. The proposed ACS (acquisition, content, and style) Taxonomy characterizes the three families of shortcuts, and the paper analyzes how simplicity bias leads to the collapse of real-fake discrimination. The paper further proposes RAIN, a multi-source few-shot AIGI detection method, to tackle shortcut learning. The main motivation is that since real images are generator-agnostic, a shortcut-free detector must judge them identically regardless of the paired generative source. Experiments validate that RAIN can achieve strong performance on unseen generators or domains.

Critical strengths
  1. The proposed method novelly treats real images as an invariance reference across source-specific heads to mitigate shortcut learning and improve generalization. It utilizes the intrinsic structure of the multi-source AIGI detection task.
Critical weaknesses
  1. There is a lack of direct empirical evidence for the main motivation of RAIN, i.e., "All heads are trained on the same real images and differ only in their fake counterpart, so their disagreement on a real image is precisely the drift that we identify as a shortcut" (Lines 196-198), despite the effectiveness of RAIN across datasets and the ablation results. This argument remains a hypothesis without sufficient evidence, and may weaken the contribution of the paper.
Questions for authors
  1. In the description of Figure 2(c), it is stated that "the contaminated detector collapses in the first epoch and then merely overfits, never learning to discriminate real from fake." However, the figure shows that the test accuracy of the "species shortcut" model increases steadily after the initial drop, which may suggest that this model actually learns some authentic real-fake discrimination afterward. How to explain this phenomenon?
3 comments

Dear Reviewer #GU5W, we would like to express our heartfelt gratitude for your precious time and expertise in reviewing our paper. We are deeply encouraged by your positive recognition of our novelty and multi-source structure. We sincerely hope the following clarifications and new analyses can address your concerns.


W1: Direct Empirical Evidence for the Disagreement–Shortcut Link.

A1: Thank you very much for your careful reading and insightful comments! We hope the following clarifications and additional results can address your concern:

  • We would like to clarify this misunderstanding that caused by our human writing. Recall that we define authenticity evidence as one-sided, such as PRNU in real images and generator fingerprints in fake images, whereas the shortcut is a two-sided attribute. Our argument is that if the heads detect forgeries based on authenticity evidence, they produce similar responses to the same real image and give largely consistent verdicts. If a large-variance disagreement occurs, it may result from shortcut reliance learned in a specific environment.
  • Following your valuable suggestion,we conduct a verification experiment based on the shortcut operators of Sec. 3. We construct 10 real–fake paired environments that share the same real images, 3 of which contain ACS shortcuts, and then train RAIN and ERM on the same environments. We evaluate both models on held-out real images that are either clean or carry one of the three attributes. We report three metrics: the disagreement rate is the fraction of unseen real images on which the heads' predictions are inconsistent. e⋆e^\star is the fraction of unseen real images for which the head with the most severe misjudgment belongs to the shortcut-reliant heads. AUC is measured with the shortcut–label association reversed at test time, so higher AUC indicates weaker shortcut reliance.
Unseen realDisagreement ↓Disagreement ↓e⋆e^\star ↓e⋆e^\star ↓AUC ↑AUC ↑
ERMRAINERMRAINERMRAIN
Clean0.2780.0030.2780.0190.8380.982
+ Acquisition0.7830.0230.7830.1270.3520.979
+ Content0.9260.0360.9260.1950.3360.878
+ Style0.5830.0080.5830.1130.3770.926

From these results, we draw two key insights:

  • Large disagreement on real images arises from shortcut reliance. Under ERM, attribute-carrying real images cause far more disagreement than clean ones, and the most severe misjudgment always comes from the three shortcut-reliant heads. Meanwhile, the AUC falls below chance under reversal. This indicates that the heads disagree exactly where some of them rely on a shortcut.

  • Cross-head consistency mitigates this reliance. RAIN reduces the disagreement and keeps a high AUC under reversal, without knowing which environments are contaminated.

We have uploaded the detailed results and the disagreement maps of real-image cases to the anonymous repository.

Q1: Accuracy Recovery of the Shortcut-Contaminated Detector in Figure 2(c), which may suggest that it learns some authentic real-fake discrimination afterward. How to explain this phenomenon?

A2: The phenomenon in Figure 2(c) validates the simplicity bias and gradient starvation in AIGI detection. We extend the training in Fig. 2(c) to 40 epochs and find that the accuracy increases steadily after the initial drop and reaches a plateau around epoch 25. We further run three independent experiments, one for each ACS shortcut, and the accuracy always remains below 0.5. Simple and easy-to-fit features produce steep loss gradients, so the model exploits the shortcut first and suppresses the learning of real–fake discrimination, which causes the initial drop on the test distribution without this shortcut. Although the accuracy then rises slowly, it stays below random guessing. This may be because not all real–fake pairs in the training data contain the same shortcut, and the shortcut strength also varies. The detector does not recover its real–fake discrimination ability; only part of its shortcut reliance is alleviated as training proceeds. We provide the new experimental curves in the anonymous repository.

Thank you for the informative response, which has addressed my concerns. Here are my follow-up comments:

  • I think A1 turns the weakness I pointed out in the original review into a strength. The motivation is now well supported by the supplemented results. From my point of view, the one-sided evidence vs. two-sided shortcut formulation is insightful and could be highlighted more in the text, and the shortcut-induced disagreement on real images is a significant observation for this task.
  • The explanations in A2 are reasonable. My guess is that the shortcut does not provide a uniformly strong, easy-to-learn feature across all training samples, so the model needs to learn other useful features, possibly including some true evidence, to further reduce the training loss. Hence, we may conclude that the model still heavily relies on shortcuts, but may not say that it does not learn any authentic real-fake discrimination/evidence.

Considering all the reviews and discussions in the rebuttal, I raise the rating from 3 to 4.

Dear Reviewer #GU5W,

We are pleased that your concerns have been addressed. We would like to express our heartfelt gratitude for your insightful suggestions and kind support! It encourages us a lot!

Best Regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.