acceptodds
Withdrawn from ICLR 2027

The Evidence Within: Tracing and Recovering Generative Evidence for Generalizable AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image detection has become increasingly important as synthetic images rapidly enter real information ecosystems. Recent visual foundation models (VFMs) equipped with lightweight linear heads have emerged as strong baselines for this task. However, like many previous detectors, they still suffer from fake-side generalization failures under unseen or shifted generators, including cross-mechanism transfer, reconstruction-style generation, and localized editing. We investigate these failures from a representation perspective: when the final image-level prediction fails, has generative evidence disappeared from the model, or does it remain encoded elsewhere in its internal representations? To study this question, we propose Mechanism-Aligned Evidence Probing (MAEP), a role-adaptive diagnostic framework that turns mechanism-specific linear heads into evidence probes and traces generation-related responses across representation depth and token types in frozen DINOv3. MAEP reveals that generative evidence is not uniformly expressed in a single representation. Its strength and persistence vary across generation mechanisms, layers, and token types: images predicted as real by the final image-level decision can still retain measurable fake-side evidence in intermediate representations or patch tokens, and the most discriminative evidence often emerges before the final layer. Motivated by this finding, we propose MIRAGE, Multi-dimensional Internal Rarity Aggregation for Generalizable Evidence, which reuses a trained linear head as an internal evidence probe to recover generation traces overlooked by the image-level decision. MIRAGE measures fake-side evidence across selected internal dimensions, converts it into empirical rarity with respect to clean-real calibration data, and aggregates complementary rarity signals through calibrated empirical min-p. Without additional backbone training, MIRAGE consistently improves generalization across diverse benchmarks and challenging distribution shifts, outperforming all compared methods on major benchmarks. In particular, it improves average balanced accuracy from 81.8% to 97.3% on AIGCDetectionBenchmark and from 91.4% to 94.3% on Chameleon, while yielding substantial gains on localized and reconstruction-style fakes. These results suggest that generative evidence follows structured, mechanism-dependent trajectories within VFM representations, and that recovering such internal evidence provides an effective route toward more generalizable AI-generated image detection. The code will be made publicly available.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
AI-generated image detection, visual foundation models, representation analysis, generalization, synthetic image forensics, calibration

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_w9LJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates whether AI-generated images misclassified by a final-layer CLS classifier still contain detectable evidence in intermediate layers and patch tokens. It introduces mechanism-aligned evidence probing (MAEP), which selects mechanism-specific linear probes and traces their responses across a frozen DINOv3 backbone. Based on these observations, the authors propose MIRAGE, a post-training method that reuses the detector’s linear head to score internal patch features. These scores are converted into empirical tail probabilities using real-image calibration data and aggregated to reconsider samples initially classified as real. Experiments report improvements across cross-generator, localized-editing, reconstruction, and degradation benchmarks.

Critical strengths

The paper addresses a concrete limitation of image-level detection: local generation traces may remain detectable even when the final CLS classifier misses them. Connecting internal representation analysis to a practical correction mechanism gives the work a clear motivation.

MIRAGE requires no additional trainable parameters beyond the existing detector. Reusing the same linear head across internal layers makes the correction procedure relatively simple, although the underlying backbone remains large.

The paper includes meaningful controls against final-CLS threshold adjustment and internal CLS-only scoring. The patch-based method outperforms these alternatives, providing evidence that the gains are not explained solely by changing the final classification threshold.

Critical weaknesses
  1. The Chameleon baseline is inconsistent across experiments. The main paper and Table 13 report DINOv3-Linear at 93.30% real accuracy, 89.54% fake accuracy, and 91.42% balanced accuracy. Table 12(a), however, reports a “matched” clean baseline at 97.34%, 87.57%, and 92.46%, respectively, while the MIRAGE result remains unchanged. Relative to the matched baseline, the balanced-accuracy improvement is 1.82 points rather than 2.86 points, and real accuracy decreases by 5.35 points rather than 1.31 points. These differences materially affect the claimed gain and its cost. The paper needs a consistent checkpoint and evaluation protocol for the central comparison. I think the result is wrong.

2.The calibration procedure establishes an empirical operating point, but its out-of-sample false-positive behavior remains insufficiently characterized. The same candidate-real reference set is used to estimate the marginal tail probabilities and calibrate their minimum. The calibration images also come from the training split used to fit the linear head; a disjoint calibration partition is not established. Matching approximately 1% FPR on these samples therefore does not demonstrate comparable control on unseen real images. At the nominal 1% setting, MIRAGE’s real accuracy is 91.99% on Chameleon and 83.06% on the original RRDataset condition. The authors acknowledge domain shift, but these substantial deviations deserve more emphasis and evaluation.

3.MAEP does not fully establish that its trajectories measure the preservation or disappearance of generative evidence. The probes are transferred across layers and token types, and some are trained on LoRA-adapted features before being applied to the frozen backbone. Changes in their margins may reflect probe–representation compatibility rather than changes in available forensic information. Probe roles are also selected manually through visual inspection, without a reproducible selection criterion. Layer-specific probes, quantitative localization results, and analysis restricted to final-CLS false negatives would strengthen the proposed interpretation.

4.The contribution of joint rarity calibration is not cleanly isolated. Table 13 compares empirical min-p against naive OR and AND rules with substantially different realized calibration FPRs: approximately 1.001%, 2.103%, and 0.376%. Their accuracy differences therefore mix aggregation choices with different operating points. Moreover, thresholding the minimum marginal p-value is itself an OR rule with a jointly calibrated cutoff. The relevant comparison is against alternative aggregation scores calibrated to the same false-positive budget.

5.Important failures remain in the scenarios central to the paper’s motivation. On DDA-COCO, MIRAGE reaches only 24.33% fake accuracy; the FLUX.1 and SD3.5-L splits remain at 5.89% and 4.14%. On DRCT-inpaint, fake accuracy is 37.97%. These are improvements over the baseline, but most fake images remain undetected. The evidence supports partial recovery rather than a broad resolution of reconstruction and localized-editing failures.

6.Generality and practical cost are not established. The main experiments use DINOv3-ViT-7B/16, with no evidence that the same approach transfers reliably to smaller backbones or other representation families. No latency, throughput, or peak-memory comparison quantifies the cost of extracting and scoring intermediate patch features. Appendix F also states that the release excludes trained heads and full data-preparation and feature-extraction pipelines, leaving substantial work required for end-to-end reproduction.

Questions for authors

1.Why do the main Chameleon comparison and Table 12(a) use different DINOv3-Linear results? Which baseline uses exactly the same head, preprocessing, and samples as MIRAGE?

2.Are the real images used for calibration disjoint from those used to train and select the linear head? How do the results change with separate reference and threshold-calibration splits?

3.What statistical interpretation is intended for the nominal 1% tolerance: an empirical calibration-set budget or an out-of-sample FPR target? Can you report uncertainty intervals and results on held-out real domains?

4.Do the MAEP trajectories remain consistent when using independently trained layer-specific probes and a fixed, quantitative probe-selection rule? Can you quantify patch localization specifically on images misclassified by the final CLS detector?

5.How does empirical min-p compare with maximum-margin, mean-score, and other aggregation rules when all are calibrated to the same realized FPR?

6.Which development data determined the layer set, K, and alpha before testing? The paper states that settings were fixed before testing but also highlights that alpha = 1% maximizes the six-benchmark mean; please clarify the selection procedure.

7.What explains the very low fake accuracy on FLUX.1 and SD3.5-L reconstruction samples? Is useful internal evidence absent, or does the current probe and aggregation procedure fail to recover it?

8.Can you provide results on a smaller backbone, measure inference overhead relative to DINOv3-Linear, and release the trained heads and scripts needed to reproduce the main tables?

Reviewer_poV2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether failures of final-layer CLS classification imply that generation evidence has disappeared from a visual foundation model. The authors introduce MAEP, which uses linear probes associated with different generation mechanisms to analyze CLS and patch representations across layers. Their observations suggest that intermediate layers and local representations may retain additional discriminative information. Building on these findings, MIRAGE reuses the base classification head and measures the rarity of internal patch responses relative to a real-image reference distribution to reconsider samples initially predicted as real. Experiments demonstrate substantial improvements across multiple benchmarks, including reconstruction and localized editing scenarios.

Critical strengths
  1. Distinguishing failures of the final classifier from the disappearance of internal discriminative information helps clarify the generalization limitations of existing detectors and motivates further use of pretrained models’ internal representations.
  2. MIRAGE reuses a single base classification head without training an additional recovery module, while using real-image calibration to constrain additional fake predictions.
  3. The evaluation covers diverse generation mechanisms, reconstruction, localized editing, and image degradation. Comparisons against direct recalibration of the final classification threshold, together with improvements in AUROC, provide evidence that internal representations offer additional discriminative information.
Critical weaknesses
  1. The authors state that the only systematic difference between the controlled real and fake subsets is whether a particular generation mechanism has been applied. However, the construction procedures remain insufficiently detailed, particularly regarding the sources, content matching, and processing pipelines of GAN and AR samples. Matching preprocessing and postprocessing alone does not rule out confounds involving image content, data sources, or processing history. Consequently, the observed separability cannot yet be clearly attributed to the generation mechanism itself.
  2. The same linear head is applied to CLS and patch features from different layers. However, equal feature dimensionality does not ensure that the classification direction has the same meaning across these representations. Changes in probe scores may reflect both changes in generation-related information and differences in compatibility between the features and the probe. The reported real–fake separation analyses and normalization provide some support, but more direct controls are needed to substantiate claims about where evidence persists or weakens. For example, independently trained and validated probes at each layer could help establish whether these conclusions remain stable.
  3. MAEP uses probes trained and selected for different generation mechanisms, whereas MIRAGE reuses a single base classification head. Information being decodable by a mechanism-specific probe does not directly establish that the base head exploits the same cues. Moreover, response trajectories over entire real and fake datasets do not sufficiently characterize the recoverable information in samples actually missed by the base detector. More direct analysis of these false negatives, using the classification head deployed by MIRAGE, would strengthen this connection.
  4. Unusual authentic content, imaging devices, or processing pipelines may also produce large internal responses. Selecting the highest-scoring patch and aggregating through the minimum tail probability allows an individual anomalous response to trigger an image-level reversal. For example, despite a calibration target of approximately 1% on training real images, the actual false positive rate on Chameleon is approximately 8.01%, compared with 6.70% for the base detector. This reflects domain shift affecting the base detector as well as the limitations of transferring a training-domain budget to unseen real domains. The authors acknowledge this limitation, but a more systematic evaluation of recovery gains against additional false positives is needed.
  5. MAEP assigns probe roles through manual inspection of response strength and persistence across layers, while MIRAGE uses a particular layer subset and patch aggregation configuration. Although the authors state that the final configuration was fixed before testing, the selection criteria, independent validation procedure, and stability across random seeds, generators, and real-image sources remain unclear. This makes it difficult to assess how strongly the conclusions depend on the particular probes and configurations selected.
Questions for authors
  1. Could you detail how the real and fake datasets for each generation mechanism were constructed, and which content-matching or processing controls rule out systematic differences unrelated to the generation mechanism?
  2. Could you analyze cross-layer and patch-level discriminative information specifically for fake images missed by the base detector, using the base head itself, and provide independently trained layer-wise probe controls to strengthen the connection between MAEP and MIRAGE?
  3. Could you clarify the independent validation procedure used to select probes, layers, and hyperparameters, and provide further calibration evaluations across real-image domains and detection results at fixed false positive rates to assess the stability of the recovery gains?
2 comments
AuthorsOpenReview ↗

Update: Main linear-head weights released

We have now released the trained main linear head H0 in the supplementary materials at supplementary_material/MIRAGE_main_H0_single_head.zip. The package includes the weights, loading instructions, and identity verification information.

Thank you for the additional experiments. The independent probes strengthen the layer-wise analysis, and the recovery results make MIRAGE's empirical gains more convincing. I am maintaining my rating of 2 (Weak rejection) for two reasons:

  1. Attribution of the observed signal. For GAN and AR, prompt matching improves the comparability of the real and generated subsets, but it does not rule out differences in image content or source. The results establish real–fake separability on the constructed datasets, but they do not substantiate Section 3's claim that applying the generation mechanism is the only systematic difference.

  2. Connection between MAEP and MIRAGE. The recovery results show that MIRAGE corrects genuine false negatives. What remains unclear is whether the deployed head uses the same cues that MAEP identifies. Comparing layer and patch responses for recovered fakes, unrecovered fakes, and correctly classified real images with similar final-layer CLS scores would help establish which signals contribute to recovery.

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

This paper aims to address the fake-side generalization failures of previous AI-generated image detectors. The proposed mechanism-aligned evidence probing (MAEP) suggests that relying solely on the final-layer CLS representation of DINOv3 can miss useful fake-side evidence in patch tokens and intermediate layers. The proposed MIRAGE method therefore calibrates the candidate "real" predictions of the final-CLS predictor with intermediate-layer and patch-token evidence. The calibration is based on real images and an empirical rarity metric. Experiments suggest the effectiveness of MIRAGE across datasets.

Critical strengths

None

Critical weaknesses
  1. There is a significant lack of formality and clarity throughout the paper. Some critical points include:
  • The "generation mechanism" is not clearly defined. It may refer to a single generator or a type of generative method. If the definition is the latter, it is questionable whether different generators of the same "mechanism" share the same characteristics or detection evidence.
  • The key questions Q1 and Q2 in Sec. 3 are informal and ambiguous.
  • The selection method of the "pre-selected dimension set" (Line 286) is not introduced, even though the selected configuration is given in the Appendix (Lines 880-882). If the selection is based on test set performance, then the method is suspicious of being affected by test data leakage.
  • The motivation and meaning of the empirical "rarity" metric are not clarified. It is difficult to understand why the comparison of fake-real margins determines some "rarity".
  • MIRAGE-Aggr at Line 377 is undefined.
Questions for authors
  1. What is the meaning of the "rarity" in Sec. 4?
  2. How is the evidence dimension set selected?
2 comments
AuthorsOpenReview ↗

Update: Main linear-head weights released

We have now released the trained main linear head H0 in the supplementary materials at supplementary_material/MIRAGE_main_H0_single_head.zip. The package includes the weights, loading instructions, and identity verification information.

Dear Reviewer, The authors have provided their responses to the raised concerns and revisions. Please review their responses and join the discussion before the end of the author-review discussion phase (Nov 18, AoE). Thanks.

Best, AC

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

Related papers

Discussion (0)

Sign in to comment.