Pith. sign in

REVIEW 3 major objections 5 minor 39 references

A training-only macro-structural feature, discarded at inference, improves AI-image detectors' cross-generator accuracy by up to 8.04 percent.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 05:48 UTC pith:JW7IYMBA

load-bearing objection A training-only SCS prior gives real cross-generator gains on some benchmarks, but the central domain-invariance claim is asserted, not shown — worth reviewing, not citable yet. the 3 major comments →

arxiv 2607.22087 v1 pith:JW7IYMBA submitted 2026-07-24 cs.CV

FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection

Md Redwanul Haque , Manzur Murshed , Manoranjan Paul , Tsz-Kwan Lee This is my paper
classification cs.CV
keywords AI-generated image detectioncross-generator generalizationimplicit regularizationScene Composition StructureLearning Using Privileged Informationshortcut learningdomain invariancezero-shot transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

F AIR (Feature-Augmented Implicit Regularization) claims that AI-generated image detectors fail on unseen generators because they overfit to generator-specific textural fingerprints. The paper's fix is to concatenate a domain-invariant 'Scene Composition Structure' (SCS) feature to the detector's latent features strictly during training, then slice it away at deployment. The resulting weight slice yields a smoother, better-calibrated decision boundary without changing the architecture or inference cost. On GenImage, FAIR raises PatchCraft's mean cross-domain accuracy by 8.04 points and AIDE's by 4.13 points; similar gains hold across four other benchmarks.

Core claim

The central claim: generalization in AIGC detection is a geometric problem, not a parameter-count problem. By tilting the separating hyperplane into an orthogonal, class-agnostic structural coordinate system (SCS) during training, FAIR forces the optimizer to discover a low-complexity boundary in the primary feature space. The prior is discarded at inference, and the sliced weights outperform the unregularized baseline on unseen diffusion and GAN domains. The paper argues that the prior's two properties — near-zero dependence on the label (I(X*;Y)≈0) and near-invariance across source and target domains (P(X*|D_source)≈P(X*|D_target)) — are what prevent shortcut learning while providing a sta

What carries the argument

The key mechanism is the FAIR training wrapper, a LUPI-inspired augmentation: the final classification layer is trained on the concatenation [X, X*], where X is the base detector's feature vector and X* is the SCS prior projected by a learned GELU MLP. At export, the slice W_X is extracted and X* is never computed again. SCS itself is computed by recursive axis-parallel partitioning of the image into segments that maximize SSE reduction; the normalized cumulative gains form the N-dimensional prior. The paper's mechanistic evidence includes weight-norm compression (‖W‖₂ drops from 24.95 to 18.42), reduced weight-update magnitudes, lower target-domain loss, and suppression of overconfident mis

Load-bearing premise

The whole argument rests on the claim that the SCS feature distribution is essentially the same for every generator, real and fake — supported only by a UMAP projection and CKA scores computed on training data; if SCS shifts across unseen generators, the invariant anchor disappears and the regularization argument collapses.

What would settle it

Compute SCS features for many generators (e.g., all GenImage domains plus unseen ones) and train a small classifier to predict the generating model from SCS. If held-out generator identity is recoverable with accuracy well above chance, the prior is not domain-invariant. Alternatively, measure the maximum mean discrepancy (MMD) between SCS distributions of source and target sets.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If FAIR is correct, any AIGC detector can be upgraded for cross-generator robustness by a training-only wrapper, with no architecture change or inference cost.
  • AIDE+FAIR reaches 91.01% mean accuracy on GenImage, the best reported on that benchmark, and PC+FAIR gains up to 18.55 points under mild JPEG compression.
  • The ablation implies that conventional L1/L2/Dropout regularization cannot produce these gains; the structural anchor is doing the work.
  • FAIR's effect is architecture-agnostic — it helps both a CLIP/DCT hybrid (AIDE) and a texture-patch model (PatchCraft).
  • The method preserves performance on source domains while improving targets, indicating reduced rather than catastrophic forgetting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: replace SCS with other macro-structural priors (depth maps, segmentation layouts) and check whether the same invariance holds; the paper's own Sobel/Laplacian ablation suggests local edge priors fail, but global layout priors may succeed.
  • The invariance premise implies a sharp falsification test: measure the SCS distribution shift between a large set of generators; if a linear classifier can recover the generator identity from SCS, the anchor is weaker than claimed.
  • If the mechanism is as general as argued, FAIR could regularize other texture-biased classification tasks (medical imaging, deepfake video) where a domain-invariant macro-structure exists.
  • The paper itself notes one failure mode: PatchCraft+FAIR drops below baseline at JPEG quality 50, where severe blocking artifacts deceive the structural anchor.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes FAIR (Feature-Augmented Implicit Regularization), a training-time method that concatenates a hand-crafted Scene Composition Structure (SCS) feature vector to the penultimate features of an AI-generated-image detector, trains the enlarged classification head, and then discards the SCS branch before deployment by slicing the weight matrix. The authors argue that SCS is domain-invariant and label-independent, so the optimizer learns a 'tilted' low-complexity decision boundary that projects back to a smoothed, generalizing boundary in the original feature space. FAIR is applied to two recent detectors, AIDE and PatchCraft, and evaluated on GenImage, AIGCDetect, UnivFD, Fake2M, and DRCT-2M. The headline results are large gains on GenImage (PatchCraft +8.04 mean accuracy, AIDE +4.13) and consistent gains on DRCT-2M and UnivFD, with some degradations on Fake2M and AIGCDetect.

Significance. If the claimed effect is real, FAIR would be a valuable and cheap addition to AIGC detection: it improves cross-generator transfer with zero inference-time overhead and without changing the deployed architecture. The manuscript has several strengths: it tests two architecturally distinct detectors, covers a large number of target domains, includes a PyTorch implementation, and compares against traditional regularizers and alternative feature priors. The core idea of using a macro-structural prior in a Learning-Using-Privileged-Information style is interesting and the empirical gains on GenImage and DRCT-2M are substantial. However, the central mechanism rests on an unverified domain-invariance assumption, and the reported comparisons are confounded by an unequal training protocol; these issues prevent the paper from being accepted in its current form.

major comments (3)
  1. [§3.3 (FAIR Implementation), Tables 1 and 5] The comparison changes two variables at once. The text states that 'the base model's primary feature extractor remains entirely frozen' during FAIR training, so only the expanded classification head (and the SCS projector) is trained. The 'Base' results appear to come from the standard end-to-end training of the full detector. Freezing the backbone and training only the head (linear probing) is itself a well-known and often strong regularizer for domain shift. The paper does not include the necessary control: the same head-only training on the frozen features without SCS, or with a random/constant auxiliary input. Without this control, the large improvements in Table 1 and the ablation in Table 5 cannot be attributed specifically to the SCS prior. Please add this ablation for both base detectors and both training regimes.
  2. [§3.4, Fig. 4, Tab. 4] The domain-invariance claim is load-bearing but not measured. The mechanism assumes P(X*|D_source) ≈ P(X*|D_target), yet the only quantitative evidence in Tab. 4 is CKA between SCS features and the real/fake label on the source training sets. CKA with the label is a test of label dependence, not a two-sample comparison of SCS distributions across generators. Fig. 4b is a qualitative UMAP projection. If SCS shifts systematically across generators (e.g., SD vs. BigGAN), the 'invariant anchor' does not exist and the tilting explanation in Fig. 1 collapses. Please provide a direct quantitative cross-domain invariance test—for example, MMD, Wasserstein distance, or A-distance between SCS distributions from the source and several held-out generators—and report the numbers.
  3. [Table 1; Tables S9, S10] The claim that FAIR 'consistently improves' cross-generator generalization is contradicted by the paper's own aggregate numbers. In Table 1, PatchCraft+FAIR degrades on Fake2M (80.46→78.82) and AIDE+FAIR degrades on AIGCDetect (93.02→92.14). No error bars or multiple seeds are reported anywhere, so for the small aggregate gains (e.g., +0.77 for PatchCraft on UnivFD) it is impossible to tell whether the difference is significant. The headline 'up to 8.04%' is the single largest improvement and is not representative of the overall behavior. Please report mean±std over at least three seeds for the aggregate and per-domain results, and temper the consistency claim accordingly.
minor comments (5)
  1. [Abstract vs. Section 1] The abstract reports 'up to 8.04%' while the contributions list says 'up to 8.10%'. Please reconcile.
  2. [Table 4] The header 'Lower is more invariant' is misleading: the reported CKA is with the label, which is a label-dependence score, not a domain-invariance measure. Rename the column and explain what the score is actually measuring.
  3. [§5.3, Tab. 4] The text equates CKA with I(X*;Y) and calls it a 'non-parametric measure' of 'conditional dependence'. CKA is a representation similarity index and is not a mutual-information estimator; the statement is technically imprecise. Please either define the actual estimator used or rephrase.
  4. [§4 / Table S7] The projection dimensions (N,M) vary across configurations without justification, and no sensitivity analysis is reported. Since M controls the capacity of the auxiliary branch, a small sweep around the chosen values would strengthen the claim that the method is robust to this hyperparameter.
  5. [Fig. 1 and §3.4] The 'tilting' and 'projection' language is illustrative. Consider labeling Fig. 1 as a conceptual schematic and moving the formal description of the weight slicing to a numbered equation for clarity.

Circularity Check

0 steps flagged

No derivation-level circularity: FAIR's gains are measured on held-out target benchmarks; SCS is a fixed pre-existing descriptor, and the domain-invariance premise is under-evidenced but not an input-output equivalence.

full rationale

The paper's headline numbers are experimentally measured on unseen target domains (Tables 1-3), not fitted values, and the FAIR training procedure (Sec. 3.3) is defined independently of those measurements. SCS is a deterministic, parameter-free descriptor introduced in the authors' prior work [12] and is not learned from the classification objective, so using it as an auxiliary training feature is not a renamed version of the target labels or of the base detector's output. The central weakness is that the invariant-anchor premise P(X*|D_source) ≈ P(X*|D_target) (Sec. 3.4) is supported only by qualitative UMAP projections (Fig. 4b) and by Tab. 4, which actually measures CKA between SCS and the label Y on source training sets; that measures label dependence, not cross-domain distribution shift. This is an evidentiary gap in the mechanistic explanation, not a circular reduction: the paper's equations do not make the benchmark improvements follow by construction, and the ablation in Tab. 5 shows that substituting non-structural priors (Random, LPIPS, Sobel, Laplacian) does not reproduce the gains. The only self-citation, [12], supplies the SCS descriptor itself and is not invoked as a uniqueness theorem or as proof of generalization. Therefore no derivation-level circularity is present; the score reflects the mild self-referential provenance of the core feature and the unverified domain-invariance claim, both of which are limitations but not circular steps.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The quantitative claim is empirical rather than derived. It depends on an unquantified domain-invariance assumption for SCS, a low-label-dependence claim measured only on source training sets, and a conceptual geometric explanation with no formal proof. The only hand-set hyperparameters are the SCS and projected dimensions, chosen per benchmark without sensitivity analysis.

free parameters (2)
  • SCS raw dimension N = 1024 or 256, depending on benchmark
    Table S7 adapts N per base method and training benchmark; no sensitivity analysis is given.
  • Projected prior dimension M = 256, 64, or 32
    Table S7 hand-sets M per configuration; no ablation across M values is reported.
axioms (4)
  • domain assumption SCS manifold is domain-invariant: P(X*|D_source) ≈ P(X*|D_target)
    Invoked in Section 3.4; supported only by qualitative UMAP plots and source-side CKA, not by quantitative target-domain statistics.
  • domain assumption SCS features have low label dependence, I(X*;Y) ≈ 0
    Used to argue the prior cannot be a shortcut; measured on the SDv1.4 and ProGAN training sets (Tab. 4), not on unseen target distributions.
  • standard math Recursive SSE partitioning greedily captures the most salient structural boundaries of an image
    Algorithm described in Section 3.1; treated as a fixed deterministic feature extractor with no optimality claim.
  • ad hoc to paper Adding an extra coordinate allows the optimizer to tilt a lower-complexity separating hyperplane, and slicing the weights back at inference preserves this smoothing
    Conceptual mechanism in Section 3.4 and Figure 1; no theorem, bound, or formal derivation connects the tilting story to the observed accuracy changes.
invented entities (1)
  • Scene Composition Structure (SCS) feature vector X* no independent evidence
    purpose: Training-only structural prior concatenated with primary features to regularize the classifier head
    SCS was published by the same authors in ref [12], but this paper provides no independent validation of its domain-invariance property beyond its own UMAP/CKA figures.

pith-pipeline@v1.3.0-alltime-deepseek · 16534 in / 11865 out tokens · 130211 ms · 2026-08-01T05:48:39.306623+00:00 · methodology

0 comments
read the original abstract

Generalization remains a critical bottleneck in AI-generated image detection. Because many modern generators are proprietary or adversarially modified, existing detectors overfit to the low-level textural patterns of accessible training data, resulting in severe failures on unseen domains. Conventional regularization techniques (e.g., $L_1$/$L_2$ norms, Dropout) apply indiscriminate parametric constraints and fail to provide the domain-invariant structure necessary for cross-generator robustness. To address this, we propose Feature-Augmented Implicit Regularization (FAIR). FAIR introduces an orthogonal, macro-structural prior, specifically, Scene Composition Structure (SCS), during training to geometrically constrain the model's optimization trajectory. By augmenting the primary feature space with domain-invariant SCS features, FAIR explicitly penalizes texture-biased shortcut learning. Crucially, this structural prior is entirely discarded at inference, yielding a smoothed, generalized decision boundary with zero architectural or computational overhead. Extensive evaluations across five massive benchmarks demonstrate that integrating FAIR into state-of-the-art detectors significantly improves cross-generator generalization, boosting accuracy by up to 8.04% and establishing new state-of-the-art robustness in zero-shot transfer scenarios.

Figures

Figures reproduced from arXiv: 2607.22087 by Manoranjan Paul, Manzur Murshed, Md Redwanul Haque, Tsz-Kwan Lee.

Figure 1
Figure 1. Figure 1: Conceptual Mechanism of Feature-Augmented Implicit Regulariza [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SCS extraction through hierarchical partitioning: Illustrating the initial cuts made by the algorithm and their resulting binary partition tree. where pi is the feature vector of the i-th pixel, and µS is the mean feature vector of S. The image is recursively partitioned by finding the optimal axis-parallel cut that maximally reduces the total SSE. For a segment S split into S1 and S2, this reduction (the … view at source ↗
Figure 3
Figure 3. Figure 3: FAIR Architecture Integration: The trainable Scene Composition Structure feature extractor (blue) generates the structural prior X ∗ which is concatenated with the primary features X strictly during training to regularize the classification head. 3.3 FAIR Implementation Feature-Augmented Implicit Regularization (FAIR) intervenes strictly at the final classification stage, as illustrated in [PITH_FULL_IMAG… view at source ↗
Figure 4
Figure 4. Figure 4: Empirical Properties of the Structural Prior ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Regularization Dynamics of FAIR. (a) Weight distributions of the final classification layer demonstrate that FAIR compresses the hypothesis space, eliminating extreme outlier weights. (b) Tracking the change in weight magnitude (∆∥W∥2) during training confirms that FAIR restrains parameter growth from the onset, indicating navigation of a smoother loss landscape. updates are directly proportional to the lo… view at source ↗
Figure 6
Figure 6. Figure 6: Source and target domain loss trajectories. FAIR imposes a slight source￾domain penalty to stabilize and reduce the out-of-distribution target loss, exemplifying the bias-variance tradeoff. Note: Target loss curves were computed post-hoc by evalu￾ating saved epoch checkpoints on a 5% subset of the target domains. This target data was used strictly for visualization and remained completely unseen by the opt… view at source ↗
Figure 7
Figure 7. Figure 7: Global distribution of misclassification confidence: aggregated across GenImage, FAIR suppresses the base model’s overconfident errors (p > 0.9), yielding a fundamen￾tally softer and better-calibrated decision boundary. 95 90 70 50 JPEG Quality Factor (Decreasing Quality ! ) 40 50 60 70 Accuracy (%) AIDE (Base) AIDE + FAIR PC (Base) PC + FAIR [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Robustness to JPEG Compression: accuracy decay curves under varying com￾pression levels show FAIR significantly enhances robustness at mild-to-moderate com￾pression qualities. algorithms act as low-pass filters that actively destroy the high-frequency micro￾textures, they provide a rigorous test for detectors that may be over-reliant on localized textural artifacts. To evaluate resilience, we subjected the… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    IEEE Trans- actions on ComputersC-23(1), 90–93 (Jan 1974).https://doi.org/10.1109/T- C.1974.223784

    Ahmed, N., Natarajan, T., Rao, K.R.: Discrete Cosine Transform. IEEE Trans- actions on ComputersC-23(1), 90–93 (Jan 1974).https://doi.org/10.1109/T- C.1974.223784

    arXiv 1974

  2. [2]

    Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization (2019),https://arxiv.org/abs/1907.02893

    Pith/arXiv arXiv 2019

  3. [3]

    Machine Learning79(1), 151–175 (May 2010).https://doi.org/10.1007/s10994-009-5152-4

    Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Machine Learning79(1), 151–175 (May 2010).https://doi.org/10.1007/s10994-009-5152-4

    doi:10.1007/s10994-009-5152-4 2010

  4. [4]

    In: Salakhut- dinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F

    Chen, B., Zeng, J., Yang, J., Yang, R.: DRCT: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. In: Salakhut- dinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F. (eds.) Proceedings of the 41st International Conference on Machine Learning. Proceedings of Machine Lear...

    2024

  5. [5]

    In: Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intel- ligence

    Cortes, C., Mohri, M., Rostamizadeh, A.: L2 regularization for learning kernels. In: Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intel- ligence. p. 109–116. UAI ’09, AUAI Press, Arlington, Virginia, USA (2009)

    2009

  6. [6]

    In: 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Corvi,R.,Cozzolino,D., Zingarini,G.,Siracusa, G.,Verdoliva,L.:Onthedetection of synthetic images generated by diffusion models. In: 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)

    2023

  7. [7]

    In: Proceedings of the 37th International Conference on Machine Learning

    Frank, J., Eisenhofer, T., Schönherr, L., Fischer, A., Kolossa, D., Holz, T.: Lever- aging Frequency Analysis for Deep Fake Image Recognition. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3247–3258. PMLR (Nov 2020)

    2020

  8. [8]

    IEEE Transactions on Information Forensics and Security7(3), 868–882 (2012).https: //doi.org/10.1109/TIFS.2012.2190402

    Fridrich, J., Kodovsky, J.: Rich models for steganalysis of digital images. IEEE Transactions on Information Forensics and Security7(3), 868–882 (2012).https: //doi.org/10.1109/TIFS.2012.2190402

    arXiv 2012

  9. [9]

    Jour- nal of Machine Learning Research17(59), 1–35 (2016),http://jmlr.org/papers/ v17/15-239.html

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-Adversarial Training of Neural Networks. Jour- nal of Machine Learning Research17(59), 1–35 (2016),http://jmlr.org/papers/ v17/15-239.html

    2016

  10. [10]

    Nature Machine In- telligence2(11), 665–673 (Nov 2020).https://doi.org/10.1038/s42256-020- 00257-z

    Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine In- telligence2(11), 665–673 (Nov 2020).https://doi.org/10.1038/s42256-020- 00257-z

    doi:10.1038/s42256-020- 2020

  11. [11]

    In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K

    Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K. (eds.) Advances in Neural Information Processing Systems. vol. 27. Curran Associates, Inc. (2014)

    2014

  12. [12]

    In: 2025 IEEE International Conference on Image Processing Workshops (ICIPW)

    Haque, M.R., Murshed, M., Paul, M., Lee, T.K.: A novel image similarity metric for scene composition structure. In: 2025 IEEE International Conference on Image Processing Workshops (ICIPW). pp. 446–451 (2025).https://doi.org/10.1109/ ICIPW68931.2025.11386242

    arXiv 2025

  13. [13]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (Jun 2016).https://doi.org/10.1109/CVPR.2016.90 FAIR for AI-generated Fake Image Detection 17

    doi:10.1109/cvpr.2016.90 2016

  14. [14]

    Nature Reviews Methods Primers4(1), 82 (Nov 2024).https://doi.org/10.1038/ s43586-024-00363-x

    Healy, J., McInnes, L.: Uniform manifold approximation and projection. Nature Reviews Methods Primers4(1), 82 (Nov 2024).https://doi.org/10.1038/ s43586-024-00363-x

    2024

  15. [15]

    IEEE Intelligent Systems and their Applications13(4), 18–28 (1998)

    Hearst, M., Dumais, S., Osuna, E., Platt, J., Scholkopf, B.: Support vector ma- chines. IEEE Intelligent Systems and their Applications13(4), 18–28 (1998). https://doi.org/10.1109/5254.708428

    arXiv 1998

  16. [16]

    Hendrycks, D., Gimpel, K.: Gaussian error linear units (GELUs) (2016),https: //arxiv.org/abs/1606.08415

    Pith/arXiv arXiv 2016

  17. [17]

    In: Ad- vances in Neural Information Processing Systems

    Ho, J., Jain, A., Abbeel, P.: Denoising Diffusion Probabilistic Models. In: Ad- vances in Neural Information Processing Systems. vol. 33, pp. 6840–6851. Curran Associates, Inc. (2020)

    2020

  18. [18]

    In: 2022 IEEE In- ternational Conference on Image Processing (ICIP)

    Ju, Y., Jia, S., Ke, L., Xue, H., Nagano, K., Lyu, S.: Fusing Global and Lo- cal Features for Generalized AI-Synthesized Image Detection. In: 2022 IEEE In- ternational Conference on Image Processing (ICIP). pp. 3465–3469 (Oct 2022). https://doi.org/10.1109/ICIP46576.2022.9897820

    arXiv 2022

  19. [19]

    Liu, B., Yang, F., Bi, X., Xiao, B., Li, W., Gao, X.: Detecting Generated Images byRealImages.In:ComputerVision–ECCV2022:17thEuropeanConference,Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XIV. pp. 95–110. Springer- Verlag, Berlin, Heidelberg (Oct 2022).https://doi.org/10.1007/978-3-031- 19781-9_6

    doi:10.1007/978-3-031- 2022

  20. [20]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV)

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9992– 10002. IEEE, Montreal, QC, Canada (Oct 2021).https://doi.org/10.1109/ ICCV48922.2021.00986

    arXiv 2021

  21. [22]

    In: Advances in Neural Information Processing Systems

    Lu,Z.,Huang,D.,Bai,L.,Qu,J.,Wu,C.,Liu,X.,Ouyang,W.:Seeingisnotalways believing: Benchmarking human and model perception of AI-generated images. In: Advances in Neural Information Processing Systems. vol. 36, pp. 25435–25447 (2023)

    2023

  22. [23]

    Online (2022),https://www.midjourney.com, Ac- cessed: 2025-09-24

    Midjourney, Inc.: Midjourney. Online (2022),https://www.midjourney.com, Ac- cessed: 2025-09-24

    2022

  23. [25]

    In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M

    Qian, Y., Yin, G., Sheng, L., Chen, Z., Shao, J.: Thinking in Frequency: Face Forgery Detection by Mining Frequency-Aware Clues. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision – ECCV 2020, vol. 12357, pp. 86–

    2020

  24. [26]

    In: Proceedings of the 38th International Conference on Machine Learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: Proceedings of the 38th International Conference on Machine Learning. pp. 8748–8763. PMLR (Jul 2021)

    2021

  25. [27]

    In: 2022 IEEE/CVF Conference on 18 M

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-Resolution Image Synthesis with Latent Diffusion Models. In: 2022 IEEE/CVF Conference on 18 M. R. Haque et al. Computer Vision and Pattern Recognition (CVPR). pp. 10674–10685 (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.01042

    arXiv 2022

  26. [28]

    Journal of Machine Learning Research15(56), 1929–1958 (2014),http://jmlr.org/papers/ v15/srivastava14a.html

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research15(56), 1929–1958 (2014),http://jmlr.org/papers/ v15/srivastava14a.html

    1929

  27. [29]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

    Tan, C., Liu, H., Zhao, Y., Wei, S., Gu, G., Liu, P., Wei, Y.: Rethinking the Up- Sampling Operations in CNN-Based Generative Network for Generalizable Deep- fake Detection. In: 2024 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR). pp. 28130–28139. IEEE, Seattle, WA, USA (Jun 2024). https://doi.org/10.1109/CVPR52733.2024.02657

    arXiv 2024

  28. [30]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Tan, C., Zhao, Y., Wei, S., Gu, G., Wei, Y.: Learning on Gradients: Generalized Ar- tifacts Representation for GAN-Generated Images Detection. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12105– 12114. IEEE, Vancouver, BC, Canada (Jun 2023).https://doi.org/10.1109/ CVPR52729.2023.01165

    arXiv 2023

  29. [31]

    In: Proceed- ings of the 38th International Conference on Machine Learning

    Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: Proceed- ings of the 38th International Conference on Machine Learning. pp. 10347–10357. PMLR (Jul 2021)

    2021

  30. [32]

    Neural networks22(5-6), 544–557 (2009)

    Vapnik, V., Vashist, A.: A new learning paradigm: Learning using privileged infor- mation. Neural networks22(5-6), 544–557 (2009)

    2009

  31. [33]

    Wang, S.Y., Wang, O., Zhang, R., Owens, A., Efros, A.A.: CNN-Generated Images Are Surprisingly Easy to Spot... for Now. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8692–8701. IEEE, Seattle, WA, USA (Jun 2020).https://doi.org/10.1109/CVPR42600.2020.00872

    arXiv 2020

  32. [34]

    In: 2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV)

    Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: DIRE for Diffusion-Generated Image Detection. In: 2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV). pp. 22388–22398. IEEE, Paris, France (Oct 2023).https://doi.org/10.1109/ICCV51070.2023.02051

    arXiv 2023

  33. [35]

    In: International Conference on Learning Repre- sentations

    Yan, S., Li, O., Cai, J., Hao, Y., Jiang, X., Hu, Y., Xie, W.: A Sanity Check for AI-generated Image Detection. In: International Conference on Learning Repre- sentations. vol. 2025, pp. 70702–70720 (2025)

    2025

  34. [36]

    In: 2018 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: 2018 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 586–595 (Jun 2018)

    2018

  35. [37]

    2019 IEEE International Workshop on Information Forensics and Security (WIFS) pp

    Zhang, X., Karaman, S., Chang, S.F.: Detecting and simulating artifacts in GAN fake images. 2019 IEEE International Workshop on Information Forensics and Security (WIFS) pp. 1–6 (2019)

    2019

  36. [38]

    arXiv (Mar 2023).https://doi.org/ 10.48550/arXiv.2311.12397

    Zhong, N., Xu, Y., Li, S., Qian, Z., Zhang, X.: PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection. arXiv (Mar 2023).https://doi.org/ 10.48550/arXiv.2311.12397

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.12397 2023

  37. [39]

    Zhu, M., Chen, H., Huang, M., Li, W., Hu, H., Hu, J., Wang, Y.: GenDet: Towards GoodGeneralizationsforAI-GeneratedImageDetection.arXiv(Dec2023).https: //doi.org/10.48550/arXiv.2312.08880

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.08880

  38. [40]

    In: Advances in Neural Information Processing Systems

    Zhu, M., Chen, H., Yan, Q., Huang, X., Lin, G., Li, W., Tu, Z., Hu, H., Hu, J., Wang, Y.: GenImage: A Million-Scale Benchmark for Detecting AI-Generated Image. In: Advances in Neural Information Processing Systems. vol. 36, pp. 77771– 77782 (Dec 2023) FAIR for AI-generated Fake Image Detection 19 A Implementation Details A.1 PyTorch Code To demonstrate th...

    arXiv 2023

  39. [103]

    Springer International Publishing, Cham (2020).https://doi.org/10.1007/ 978-3-030-58610-2_6

    2020