acceptodds
Withdrawn from ICLR 2027

Beneath Semantics: Signal‑axis Residual Fingerprint for AI‑Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image detectors perform well on aggregate benchmarks but exhibit marked declines in performance under distribution shift—whether from new collection pipelines for familiar generators or from entirely unseen generators. We examine image authenticity along two complementary evidence axes: semantic representations, which provide detection cues through pretrained image representations, and signal statistics, which capture local dependencies and spectral organization in residuals. To address both collection shift and unseen-generator shift, we introduce Signal-axis Residual Fingerprint (SRF), a compact detector built around a fixed representation of Spatial–Spectral Microstructure. Rather than learning dataset- or generator-specific patterns, SRF uses six fixed resampling operators and an interpretable statistic bank to characterize local dependencies and spectral organization in residuals—traces intrinsic to the generation pipeline rather than specific to a particular collection protocol or generator. The complete five-block configuration, SRF-E, contains only 0.28M fitted parameters, requires no pretrained backbone, and achieves a mean standalone AUROC of 0.9150 across 48 benchmark–generator groups from four benchmark suites. Crucially, its scalar score can be fused with the 12 evaluated frozen-backbone probes to recover performance under both distribution shifts. For the strongest CLIP probe, mean AUROC increases by approximately 2.6 percentage points, worst-group AUROC rises from 0.691 to 0.817, and mean AUROC on the unseen-generator tier improves from 0.855 to 0.899. Controlled transformation experiments show that periodic residual correlations and spectral energy statistics provide processing-related evidence that the current CLIP detection scores do not fully exploit.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, media forensics, residual statistics

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_nEcQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Signal-axis Residual Fingerprint (SRF), a lightweight detector based on fixed resampling residuals and spatial–spectral statistics. The paper also studies the fusion of SRF-E with semantic detectors such as CLIP-LP under collection shift and unseen-generator shift. The problem is relevant, and the evaluation protocol attempts to separate these two types of distribution shift. However, the experimental comparisons are not sufficiently up to date. All compared methods were published in 2025 or earlier, which is not sufficient for an ICLR 2027 submission and makes it difficult to assess the actual competitiveness of SRF-E. Moreover, the proposed fixed equal-weight fusion strategy lacks adaptability and can severely drag down performance under strong post-processing.

Critical strengths

1.This paper studies both collection shift and unseen-generator shift, instead of reporting only aggregate performance on standard benchmarks.

2SRF-E is lightweight, uses no pretrained backbone, and provides an interpretable signal-axis score that can be fused with semantic detectors.

Critical weaknesses

1 The experiments do not sufficiently compare against recent methods. All baselines in Table 1 were published in 2025 or earlier. This is not sufficient for an ICLR 2027 submission.

2.The novelty relative to NPR and prior low-level artifact detectors is not sufficiently clarified. The main difference appears to be that SRF replaces a learned feature extractor with fixed resampling operators and hand-designed statistics.

3.There is a lack of direct comparison of ACC results.

4.The robustness results do not fully support the proposed signal-axis interpretation. SRF-E is sensitive to JPEG compression and other post-processing operations, and its performance can decrease substantially under strong compression.

Questions for authors

1.The compared baselines are limited to 2025 or earlier. Why are more recent state-of-the-art methods not included in the main comparison? ( ICLR2026, CVPR2026,ICML2026, etc)

2.The main evaluation exclusively relies on AUROC and AP. Since Accuracy (ACC) is a standard and widely used metric in recent works, why is it only marginally reported in the appendix (Tab 9) rather than serving as a primary metric in the main paper?

  1. The authors reveals that fixed operators and static fusion perform poorly under heavy image degradation (e.g., JPEG-40) and strict low FPR settings. Given these severe practical vulnerabilities, why persist with a strictly fixed extraction approach instead of exploring an adaptive or reliability-aware mechanism?
Reviewer_HQesOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Signal-axis Residual Fingerprint (SRF), a detector built from six fixed downsampling–upsampling residual operators and five blocks of spatial and spectral statistics. The 5,742-dimensional descriptor is scored using shrinkage-LDA members, real-reference calibration, clean/JPEG training views, and a conditional Mahalanobis OOD guard. The default SRF-E has 0.28M fitted parameters without a pretrained backbone. Validation-standardized score fusion improves the default CLIP-LP from 0.9363 to 0.9622 in mean AUROC over 48 benchmark–generator groups, and improves the mean performance of all 12 evaluated frozen-backbone probes. The paper also studies benchmark-release shifts, unseen generators, controlled processing interventions, and label-conditioned score correlations.

Critical strengths
  1. A clear and reproducible complementarity result. Adding SRF-E to the same CLIP-LP yields +0.0259 mean AUROC over 48 groups, +0.0449 on the 14 unseen groups, and raises minimum group AUROC from 0.6912 to 0.8171. Positive gains across 12 frozen representations indicate the result is not specific to one CLIP architecture.
  2. Attractive performance/storage trade-off. 0.28M fitted parameters and a 2.5 MB checkpoint without pretraining make this appealing as an auxiliary branch; the paper appropriately limits "lightweight" to parameters and storage.
  3. Reporting transparency. Per-group scores, block and guard ablations, 17 post-processing conditions, validation-frozen operating points, low-FPR diagnostics, and deduplication audits are provided, with negative transfer, a CI crossing zero, and the low-FPR shortfall disclosed explicitly.
Critical weaknesses
  1. Processing sensitivity does not establish generation specificity. The abstract motivates the descriptors as traces intrinsic to generation rather than to a collection protocol. Yet Figure 4 shows that bicubic down/up resampling of real photographs at scale 0.75 flips 18.4% of them from real to fake under SRF-E, against 1.0% for CLIP-LP. This establishes sensitivity to that operation, not exclusivity to generated images; camera processing, demosaicing, denoising, and sharpening are alternative sources of such structure.
  2. The incremental necessity of the full descriptor design is unestablished. Table 2 attributes the largest losses to cooc (−0.0134 fused) and self (−0.0074). Removing spec or E costs only 0.0011 / 0.0004, and removing E actually raises standalone worst-group AUROC from 0.6214 to 0.6350. Missing is a matched comparison — same training sources, JPEG augmentation, readout, calibration, and guard — against simple fixed high-pass/SRM statistics, a compact cooc+self variant, and single-operator configurations. This leaves unclear whether the contribution is a new spatial–spectral construction or an effective configuration of established tools.
Questions for authors

1.Under source- and processing-matched controls, how do the gains and the false-positive rate on real images change? How is generation-related evidence separated from acquisition/post-processing cues?

Reviewer_iWntOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Signal-axis Residual Fingerprint (SRF), a lightweight framework for AI-generated image detection that operates without a pretrained backbone and complements existing semantic detectors. SRF extracts Spatial–Spectral Microstructure (SSM) from residuals between input images and fixed resampling reconstructions, using interpretable statistics and a lightweight Shrinkage-LDA classifier for detection. It also incorporates an OOD Guard to mitigate misclassification of images from unseen generators and can be integrated with existing semantic detectors without training an additional fusion module.

Experiments across multiple benchmarks demonstrate that SRF achieves competitive standalone detection performance and improves generalization under collection shifts and unseen generators when combined with semantic detectors. Evaluations across various pretrained backbones and controlled image transformation experiments further support that SRF captures low-level signal characteristics complementary to those exploited by existing semantic detectors.

Critical strengths
  1. The paper proposes a novel detection approach that leverages spatial and spectral residual statistics as cues for AI-generated image detection. By combining these cues with existing detectors using pretrained backbones, the proposed method achieves significant performance improvements across various benchmarks.
  2. The proposed method is lightweight and utilizes interpretable residual statistics and linear classifiers, offering advantages in terms of interpretability and computational efficiency.
  3. The paper provides extensive experiments across diverse benchmarks, along with detailed ablation studies and additional analyses, to demonstrate the effectiveness of the proposed method and validate the contributions of its individual components.
Critical weaknesses
  1. The manuscript is generally difficult to read, and the core ideas and design motivations of the proposed method are not clearly conveyed. Although several components, including the Shrinkage-LDA Committee, OOD Guard, and Three-tier Evaluation, are introduced, their individual roles, necessity, and relationships are not sufficiently explained. The writing is often overly complicated and mechanical, making it difficult and time-consuming to understand the overall framework and its contributions.
  2. The motivation for adopting Shrinkage-LDA to classify the proposed residual statistics is not sufficiently justified, and comparisons with alternative classifiers, such as Logistic Regression and SVM, are lacking. Furthermore, the theoretical motivation and complementary roles of the Shrinkage-LDA Committee and Mahalanobis OOD Guard remain unclear. In particular, the paper provides limited analysis of whether the OOD Guard reliably corrects misclassifications without introducing additional errors.
  3. The design and validity of the Three-tier Evaluation framework for distinguishing Collection Shift from Unseen Generator Shift are not sufficiently justified. In particular, each tier is closely associated with specific benchmark datasets, making it difficult to determine whether the observed performance differences arise from dataset-specific characteristics or generator novelty.
  4. Although Community Forensics is used as a major evaluation benchmark, the detector proposed in the original study is not included in the baseline comparisons. Given that its pretrained checkpoint is publicly available, the reason for this omission is unclear. This limits the completeness of the experimental comparisons, particularly in evaluating generalization to unseen generators.
Questions for authors
  1. Although the ablation study demonstrates benefits from the OOD Guard, the rationale for combining it with the Shrinkage-LDA Committee and their complementary roles remain unclear. Could the authors provide a more rigorous justification or analysis showing how reliably the OOD Guard corrects LDA misclassifications without introducing additional errors?
  2. Why was Shrinkage-LDA selected for classifying the proposed residual statistics? Have alternative classifiers, such as Logistic Regression, SVM, or MLP, been considered? Theoretical justification or empirical comparisons would help support this design choice.
  3. The motivation and validity of the Three-tier Evaluation could be explained more clearly. Since each tier is closely associated with specific benchmarks, dataset characteristics may be confounded with generator novelty. Could the authors clarify the tier assignment criteria and provide evidence that this protocol effectively distinguishes Collection Shift from Unseen Generator Shift?
  4. The paper uses Community Forensics as an evaluation benchmark but excludes its proposed detector [1] from the baseline comparisons, despite the availability of pretrained checkpoints. Could the authors explain this omission and, if feasible, provide additional comparisons?
  5. Although the paper includes a Reproducibility Statement, implementation code and trained checkpoints are not provided with the submission. Do the authors plan to release these resources, along with training configurations and details of the training data and Real Reference Set, to support reproducibility and independent verification?

I appreciate the authors' efforts to advance AI-generated image detection. While the proposed residual statistics approach and extensive experiments are promising, the manuscript lacks clarity in its methodological justification and overall presentation. The absence of code and checkpoints also raises reproducibility concerns. I would consider increasing my score if the authors adequately address these issues, particularly methodological justification and reproducibility.

[1] Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025)

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

Related papers

Discussion (0)

Sign in to comment.