acceptodds
Withdrawn from ICLR 2027

Source-Guided Inference-Time Rectification for More Generalizable AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Rapid advances in AI-generated imagery (AIGI) call for reliable and generalizable detectors capable of identifying increasingly diverse and realistic synthetic imagery. However, existing detectors often suffer substantial performance degradation when confronted with unseen generators, manipulation techniques, or post-processing operations. We revisit this problem from the perspective of internal representations and observe a consistent association between detection failures and test representations deviating from reliable source-domain regions, while such deviations are partially rectifiable at inference time. Motivated by these observations, we propose Source-Guided Subspace Rectification (SGSR), a plug-and-play inference-time representation rectification framework that can be integrated into existing AIGI detectors to improve their cross-domain generalization. SGSR models an observed test representation as a perturbed observation of an underlying reliable representation and rectifies it according to a reliable representation distribution estimated from correctly predicted source-domain samples, whose covariance structure enables direction-dependent rectification. SGSR requires no target statistics, pseudo-labels, cross-sample memory, or parameter updates during inference, while introducing only marginal additional computational and storage overhead. Extensive cross-dataset and cross-generator experiments on both Deepfake and synthetic image detection benchmarks demonstrate consistent improvements, with average AUC gains of 0.80 and 4.08 percentage points across 19 Deepfake detectors and 13 synthetic image detectors, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Image Detection, Cross-Domain Generalization, Inference-Time Rectification

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_fhZ6OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SGSR, an inference-time feature rectification method for AI-generated image detection. It estimates a reliable representation distribution from correctly classified source samples and uses Gaussian MAP estimation with low-rank covariance approximation to rectify test features.

Critical strengths
  1. Extensive evaluation. Experiments across 32 detectors demonstrate consistent improvements, with average AUC gains of 0.80 points for deepfake detection and 4.08 points for synthetic image detection.

  2. Practical efficiency. The method requires no retraining, target-domain statistics, or parameter updates, introducing minimal inference overhead.

Critical weaknesses
  1. Questionable modeling assumptions and limited evidence of reliability. The method assumes that correctly classified source representations characterize a reliable distribution and that domain shift behaves like isotropic Gaussian noise. However, correctly classified samples may still rely on domain-specific shortcuts, and a single Gaussian shared by real and fake samples may ignore class-discriminative structure. Empirically, the deepfake gains are small and reported without variance estimates, and the larger synthetic-image gains come mainly from weaker baselines on a few small subsets. This matters because the assumptions are not validated, and the evidence does not establish a reliable benefit for strong detectors.

  2. Limited novelty and unclear mechanism. SGSR is a fixed linear shrinkage determined only by source statistics. Applied at the last layer, it is equivalent to replacing the classification head, and it is closely related to PPCA-style denoising and existing subspace projection methods such as RFI. Moreover, improved accuracy does not show that the method recovers reliable latent representations. This matters because the central framing of sample-wise rectification is not supported, and the contribution over equivalent linear modifications is unclear.

Questions for authors
  1. Equivalence to a linear head change. When SGSR is applied only at the last layer, is it equivalent to replacing the classifier weights with a source-derived linear transformation of them? How does last-layer-only rectification compare with applying it at all layers? If most of the gain comes from the last layer, the method reduces to a fixed head modification.
  2. Comparison with closely related linear methods. Under the same detectors and protocol, how does SGSR compare with PPCA denoising, hard projection onto the source principal subspace, and RFI? This determines whether SGSR adds anything beyond existing subspace projection.
  3. Validity of the reliable-distribution assumption. How do results change with a class-conditional prior estimated separately for real and fake samples, and with statistics estimated from all source samples rather than only correctly predicted ones?
Reviewer_sSSAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Source-Guided Subspace Rectification (SGSR) to improve the generalization of AI-generated image detectors. The main motivation is that detection accuracy tends to decrease when test representations move away from correctly classified source-domain representations. Based on this, SGSR models test features as noisy observations of reliable latent representations and uses source-domain statistics to rectify them during inference. A low-rank covariance approximation makes the correction efficient and applicable to different detector architectures. Experiments on 19 Deepfake detectors and 13 synthetic image detectors show average AUC improvements.

Critical strengths
  • Practical and well-motivated problem. The paper addresses cross-domain generalization without requiring retraining or access to target-domain data. SGSR can be applied to existing CNN- and Transformer-based detectors without changing their parameters. The closed-form rectification and low-rank approximation also make the method relatively easy to integrate.
  • Extensive evaluation across detectors. The experiments cover 19 Deepfake detectors and 13 synthetic image detectors, including different architectures and adaptation strategies. The evaluation also covers different unseen datasets and generation techniques, which helps assess its generalization ability.

  • Useful empirical analysis. The paper examines the relationship between representation deviation and detection accuracy, compares different rectification formulations, and reports inference overhead. I think these experiments help explain the motivation and practical behavior of SGSR.

  • Simple and interpretable formulation. The proposed rectification has a clear probabilistic interpretation, combining the observed test representation with a source-estimated reliable representation distribution.
Critical weaknesses
  • The assumption that reliable representations follow a shared Gaussian distribution is not very clear. SGSR models correctly classified real and fake features using a single Gaussian distribution. However, these features may have different distributions. It is unclear whether this shared Gaussian assumption is appropriate or whether it could weaken the distinction between real and fake images.

  • The improvements on Deepfake detection are relatively modest. Although SGSR improves the average AUC across all 19 Deepfake detectors, the overall gain is only 0.80 percentage points. It would be useful to better characterize which detectors and domain shifts benefit most from rectification.

  • The comparison with existing inference-time methods is insufficient. The paper compares mean interpolation, full-covariance rectification, and SGSR, but these are closely related variants of the proposed approach. More comparisons with existing inference-time feature correction or adaptation methods would be helpful.

Questions for authors
  • Limited analysis of when rectification hurts performance. Table 4 shows that SGSR corrects 15.67% of wrong predictions but also changes 3.75% of correct predictions into incorrect ones. It would be useful to analyze why rectification harms these samples and whether such negative effects can be reduced.

  • Several recent studies also investigate generalization in AI-generated image detection, including Beyond Generation (CVPR 2025), Towards Generalizable Detector for Generated Image (NeurIPS 2025), and Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection (ICML 2026). These methods also consider the treatment of source distributions and unseen generators.

Reviewer_1QJvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Source-Guided Subspace Rectification (SGSR), an inference-time framework for improving the generalization of AI-generated image detectors. SGSR estimates a reference representation distribution from correctly classified source-domain samples and applies a closed-form, covariance-aware correction to test representations. A low-rank approximation of the source covariance enables efficient rectification at intermediate layers of CNN- and Transformer-based detectors. The method uses fixed source statistics without requiring target-domain statistics, pseudo-labels, or model parameter updates during inference.

Critical strengths
  1. An interesting inference-time perspective. Inference-time representation rectification is relatively underexplored in generalizable AI-generated image detection. SGSR provides a useful complement to training-based approaches by using fixed source-domain statistics without requiring target-domain statistics, pseudo-labels, or parameter updates during inference.

  2. Consistent improvements in average AUC across existing detectors. SGSR improves the average AUC of all 19 deepfake detectors in Table 1 and all 13 synthetic-image detectors in Table 2. These results across different architectures and adaptation strategies support its potential as a broadly applicable addition to existing detectors.

Critical weaknesses
  1. The main evaluation does not include accuracy. Tables 1 and 2 report AUC but do not provide corresponding accuracy results for the original detectors and SGSR. Although the appendix includes preliminary accuracy results and prediction-transition analyses, these do not replace a systematic evaluation of classification performance. Notably, Table 18 shows that Effort on DFDC corrects 1,304 previously incorrect predictions but changes 1,580 previously correct predictions to incorrect ones, despite its AUC improvement in Table 1. Reporting accuracy or balanced accuracy under a clearly specified threshold-selection protocol would clarify whether improved ranking translates into better detection decisions.

  2. Comparisons with existing inference-time interventions remain limited. Appendix Table 10 evaluates Feature Clipping (Tao et al., 2025), ReAct (Sun et al., 2021), and Robust Feature Inference (RFI; Singh et al., 2024) on four deepfake detectors but does not cover cross-generator synthetic-image detection, where SGSR achieves its largest gains. Broader comparisons under matched evaluation and hyperparameter-selection protocols would clarify when SGSR is preferable to existing interventions. I would suggest including Feature Clipping, ReAct, and RFI in Tables 1 and 2 to enable direct comparisons.

  3. The AP improvement reported in Section 4.2 lacks supporting results. Section 4.2, line 428, states that SGSR improves average precision from 93.42 to 96.53, but no corresponding AP table appears in the main paper or appendix. The stated gain of 3.10 points should also be clarified, since the displayed values differ by 3.11 points, potentially because of rounding. This unsupported statement raises concerns about the soundness of the paper’s evaluation.

References

Tao et al. (2025). Feature Clipping for Uncertainty Calibration. AAAI 2025.

Sun et al. (2021). ReAct: Out-of-Distribution Detection with Rectified Activations. NeurIPS 2021.

Singh et al. (2024). Robust Feature Inference: A Test-Time Defense Strategy Using Spectral Projections. Transactions on Machine Learning Research (TMLR).

Questions for authors
  1. What determines the magnitude of the improvement? To what extent can SGSR improve existing methods, and how does the underlying backbone architecture affect the magnitude of the improvement? For example, Table 2 shows that methods with higher AUC values tend to achieve smaller gains, which is reasonable because they have less room for improvement. However, this pattern does not appear to hold for the CNN-based forgery detectors in Table 1, whose AUC values range from 60 to 70 yet gain only 0.22–0.65 points. Could the authors explain this difference?

  2. Comparison with existing inference-time interventions. Could the authors include Feature Clipping, ReAct, and RFI in Tables 1 and 2 to enable direct comparisons across both detection settings? Feature Clipping, published in 2025, is the most recent of these three methods. Are there more recent methods that could also be evaluated? Could the authors also report accuracy alongside AUC for all compared methods, using a clearly specified threshold-selection protocol.

  3. Supporting evidence for the reported AP improvement. Section 4.2 states that SGSR improves average precision from 93.42 to 96.53, but no corresponding AP table appears in the main paper or appendix. Could the authors provide detailed AP results?

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

Related papers

Discussion (0)

Sign in to comment.