acceptodds
Under review as a conference paper at ICLR 2027

GENERATOR FINGERPRINT ISOLATION: ADDRESSING BLIND SPOTS IN AI-GENERATED IMAGE DETECTION

Abstract

AI-generated image detection is an open-set generalization problem: detectors trained on a limited set of generators may fail to generalize to unseen generators, resulting in generator-specific blind spots. A common response is to adapt the detector using samples from a missed generator, but this modifies its decision function and may affect generators it already detects reliably. We instead ask whether such blind spots can be repaired by adding generator-specific evidence while preserving the original detector. We propose Generator Fingerprint Isolation (GFI), a training-free framework that augments a frozen detector with such evidence. GFI exploits an empirical property of pretrained representations such as CLIP and DINO: the principal components that account for most variation among real-image embeddings can be removed without eliminating generator-discriminative signal. Across the blind-spot generators considered in our evaluation, removing the principal components accounting for 95% of real-image variance yields a residual space that preserves 78.8% of the measured generator-discriminative signal. GFI uses this residual space to extract a generator-specific fingerprint for each blind-spot generator. At inference, alignment with the fingerprint provides generator-specific evidence that is fused with frozen detector's prediction. Fingerprints can be registered independently as new blind-spot generators are encountered. We evaluate GFI on AIGCDetectBench, with 17 GAN, diffusion, and commercial generators, across three off-the-shelf modern detectors—DDA, Effort, and SPAI. For each detector, we identify its own blind-spot generators and repair them using generator-specific fingerprints. Across these targeted generators, GFI improves mean AUC from 0.827 to 0.972 for DDA, 0.747 to 0.914 for Effort, and 0.762 to 0.900 for SPAI, while also substantially reducing FPR@95% Recall. AUC improvements are statistically significant for every detector on its identified blind-spot generators under paired DeLong tests. For generators not targeted for repair, GFI shows no systematic degradation. These results demonstrate that generator-specific evidence can repair targeted blind spots while preserving the detector.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Image Detection, Image Forensics, Image Generator Fingerprints

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1bgKOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the open-set generalization problem in AI-generated image detection, specifically focusing on "blind spots" where detectors trained on limited generators fail on unseen ones. The authors propose Generator Fingerprint Isolation (GFI), a training-free framework that augments frozen detectors with generator-specific evidence without modifying their parameters. The key idea leverages an empirical observation: removing principal components accounting for 95% of real-image variance preserves ~78.8% of generator-discriminative signal in the residual space. GFI constructs residual projections from real images, extracts normalized mean displacement fingerprints for each target generator, calibrates alignment scores against held-out real images, and fuses this evidence with the frozen detector's output.

Critical strengths
  1. Reframing detection failures as addressable "blind spots" rather than requiring full retraining is practically valuable and well-motivated. GFI requires no gradient updates to the base detector, making it lightweight and practical.
  2. The key empirical finding that generator-discriminative signal concentrates in low-variance directions after removing dominant real-image variation is interesting.
Critical weaknesses
  1. The paper does not analyze how many images are needed for reliable fingerprint estimation, nor the sensitivity to the quality/diversity of the fingerprint set. Additionally, the "post-hoc repair" setting assumes the user has identified which generator is the blind spot, this requires additional machinery to detect failures.
  2. The novelty is limited. The main conceptual that the empirical finding about signal retention in residual space is interesting, but the paper does not provide theoretical justification or deeper analysis of why this holds.
  3. The comparison is limited to fine-tuning baselines. Given the paper's focus on AIGCDetectBench, it should compare against state-of-the-art cross-generator detection methods.
Questions for authors
  1. Could you provide a deeper analysis of why removing the top-95% of real-image variance preserves most generator-discriminative signal? Is this specific to CLIP/DINO features, or would it hold for other encoders?
  2. How many self-generated images are needed for reliable fingerprint estimation? How sensitive is performance to the specific images used in the fingerprint set?
  3. For the generators that remain below 0.90 after GFI, what specific challenges exist?
  4. For generators where fingerprint alignment alone achieves 100% AUC (Table 2), what does this imply? Are these generators trivially detectable by any method?
  5. How does a user know a generator is a "blind spot" before registering it?
4 comments

We thank the reviewer for their feedback and address each concern below.

1. Signal Retention & Encoder Generality

Pretrained vision encoders like CLIP capture semantic content such as objects, scenes and lighting. Their dominant feature directions contain substantial natural variation across real images. Generator artifacts may also occur in these directions, but are harder to distinguish against this background. Importantly, artifacts persist in lower-variance directions, where real-image variation is smaller. Removing the top 95% of real-image variance (Eqs. 4–8, App. C.1) reduces this background while retaining 78.8% of generator-discriminative signal, as quantified by Eqs. 7–8 (Table 2, Fig. 3).

Two existing analyses support this:

  • Component replacement (App. D.1): Replacing a generated image's dominant PCA coefficients with those from a random real image barely affects detection (ΔAUC < 0.001), indicating that the content captured by these dominant directions contributes little to detection.
  • Numerical stability (App. C.2): Since discriminability is measured relative to real-image variation, tiny eigenvalues could inflate discriminability estimates. Flooring small variances (10⁻¹²–10⁻⁶) leaves retained signal unchanged (78.7%).

We add two ViTs (AugReg, MAE) and use unbiased Crossnobis distances to avoid noisy overestimates. All four retain 73–89% of generator signal in 86–90% of directions:

Encoder (dirs kept)CLIP (86%)DINO (90%)AugReg (89%)MAE (88%)
Standard/Crossnobis.783/.775.893/.887.766/.733.853/.842

Together, these results show that dominant-variance removal suppresses natural content variation while retaining generator-discriminative signals across encoders.

2. Sample Size and Sensitivity

We use 20 random subsets per sample size to assess sensitivity to image selection.

DetectorBaseN=10N=50N=100
DDA (4).827.945±.042.952±.008.970±.007
Effort (3).747.904±.013.904±.008.913±.003
SPAI (4).762.846±.015.880±.004.900±.003
All 11.782.898±.024.913±.007.929±.004

Just 10 images capture over two-thirds of the improvement. At 50–100 images, gains increase with low variability (SD ≤.008), showing consistent improvements across diverse generators. We use N=100 for experiments.

3. Why Does WFIR Remain Below .90 on Effort and SPAI?

AIGCDetectBench's WFIR pairs heavily web-compressed StyleGAN faces with Flickr real faces, making detection harder. The standalone fingerprint bank (no fused detector) tested on AIGCDetecBench achieves .71–.77 AUC (Table 11). GFI fusion (Table 3) combines fingerprint evidence with the base detector: a strong base like DDA pushes WFIR above .90 (from .87), while weaker bases also improve (Effort: .75→.84; SPAI: .58→.72), but they remain below .90 threshold. Augmenting enrollment samples to improve fingerprint robustness is a natural extension (7Vjw, #3).

4. What Does 100% AUC Mean?

Table 2 measures isolated fingerprint separability: whether a generator's fingerprint distinguishes its clean outputs from held-out real images. AUC 1.00 shows that this generator-specific signal is recoverable after semantic removal, not that existing detectors detect the generator reliably (Table 1). On benchmark images, the standalone fingerprint bank achieves only .71–.77 AUC for WFIR (Table 11), showing that perfect separability in this controlled evaluation does not imply perfect detection in practice.

5. Unknown Generator Blind Spots.

GFI does not require generator identities for enrollment. We demonstrate this by clustering unlabeled images in residual CLIP space, where images with matching fingerprint directions naturally group together.

MethodEstimated groups (actual: 7)PurityNMIARI
GMM-BIC6.7±0.4.951.980.948
Eigengap7.6±0.5.985.933.903

Both methods recover generator-specific clusters: GMM-BIC estimates group count more accurately; eigengap achieves higher purity. Without isolation γ=0\gamma=0, purity drops to .857/.849, and both estimate six groups. Thus, GFI can discover generator-specific groups without knowing their identities in advance.

6. Comparison With Cross-Generator Detectors

Table 1 benchmarks seven published detectors across all AIGCDetectBench generators, revealing generator-specific blind spots across diverse approaches. GFI addresses these as a training-free repair layer for frozen detectors, rather than a standalone detector. We therefore compare it with fine-tuning, which can also repair blind spots but requires retraining. We select DDA, Effort and SPAI for their distinct detection mechanisms. Improvements across all three demonstrate GFI's applicability across detector types; it can also extend to the remaining Table 1 detectors.

Thank you again for your thoughtful review. As the discussion period nears its close, we wanted to gently follow up to check whether our response addresses your concerns. Please let us know if you have any further questions we can answer before the deadline.

Thank you for your detailed response. However, regarding Figure 1 and the writing style throughout the paper, I find that the logical progression is lacking and the tone is overly defensive. I suspect this paper may have been generated using automated research tools; therefore, I am maintaining my score but lowering my confidence rating to 3–2. I hope this assists the Area Chair in their assessment.

Thank you for taking the time to follow up. We appreciate your feedback on Figure 1 and the flow of the writing, and we will certainly work on making both clearer. At the same time, we want to respectfully push back on the suggestion that this paper was produced by automated research tools.

To share our side clearly: every part of the research—the core ideas, theoretical derivations, experimental setup, ablations, and analysis—was designed and run by us. Following the guidelines, we declared using AI only to help polish the phrasing of our draft, not to generate the work itself. Since the initial review focused entirely on the technical details, we were quite surprised by this new concern after the rebuttal. Could you share what specifically gave you that impression, or which parts of the logical progression felt disconnected? We would genuinely value understanding your reasoning.

Lastly, can you let us know whether our earlier responses answered your technical questions, or if there are any other doubts we can clarify?

Reviewer_V26POpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Generator Fingerprint Isolation (GFI), a training-free framework for repairing generator-specific blind spots in AI-generated image detectors. GFI extracts generator fingerprints from residual pretrained features, calibrates their detection scores using real images, and combines them with frozen detector predictions. Experiments across multiple detectors and generators demonstrate improvements on targeted blind spots without substantial aggregate degradation on non-target generators.

Critical strengths
  1. Practical and modular framework. GFI addresses an important deployment problem by allowing existing detectors to incorporate generator-specific evidence without modifying their parameters. The ability to register fingerprints independently is particularly appealing for maintaining deployed detection systems.

  2. Strong empirical improvements with useful diagnostic analyses. The method improves detection performance across three different base detectors and multiple generator-specific blind spots. The authors also investigate residual feature discriminability, cross-generator transfer, statistical significance, and performance on non-target generators, providing substantial empirical support for the proposed approach.

Critical weaknesses
  1. Limited methodological novelty. GFI primarily combines established techniques, including PCA-based residual projection, mean-feature prototypes, statistical calibration, and score fusion. Closely related works such as NS-Net [1] and LoRC [2] have already explored residual representations for AI-generated image detection. The paper does not sufficiently establish a substantial methodological contribution beyond integrating these existing techniques.

  2. Insufficient validation of the core design. The necessity of PCA-based fingerprint isolation is not convincingly demonstrated. Appendix D.2 shows that removing dominant components at the selected operating point actually yields lower fingerprint-only AUC than using the full feature space (0.984 vs. 0.991). Without systematic comparisons against simpler prototype-based approaches under identical calibration and fusion settings, it remains unclear whether the proposed isolation mechanism provides meaningful benefits.

  3. Limited scalability and reliability in open-set deployment. GFI requires identifying blind-spot generators and obtaining generator-specific samples, limiting its applicability to unknown or inaccessible generators. The paper also lacks an analysis of how performance and calibration scale with the number of registered fingerprints. Moreover, Table 4 reveals increased false-positive rates under real-image distribution shifts, particularly for SPAI (0.10 to 0.16). These limitations raise concerns about whether GFI can reliably extend existing detectors without compromising their performance.

References

[1] Yan, J., Wang, F., Jiang, W., Li, Z., and Fu, Z. (2025). NS-Net: Decoupling CLIP Semantic Information Through NULL-Space for Generalizable AI-Generated Image Detection. arXiv:2508.01248.

[2] Yan, H., et al. (2026). LoRC: Detecting AI-Generated Images via Low-Rank Collapse in Semantic Residuals. ECCV 2026.

Questions for authors
  1. Can the authors compare GFI against simpler alternatives, such as full-feature mean prototypes or nearest-prototype classifiers, under identical calibration and fusion procedures? What specific benefit does PCA-based fingerprint isolation provide?

  2. How does GFI behave as the number of registered fingerprints increases? Can the authors report the relationship between fingerprint-bank size, target detection performance, and false-positive rates on unseen real-image distributions?

  3. How can GFI address blind spots when the underlying generator is unknown or inaccessible? What advantages does it offer over conventional few-shot adaptation in this setting?

3 comments

We thank the reviewer for their feedback and address each concern below.

1. GFI Novelty

As generators evolve, blind spots remain a persistent challenge for AI-image detectors. While most detectors aim for broad generalization, fine-tuning for each missed generator is operationally very costly. Training-free alternatives rely on fragile pixel residuals, generator identity, or image references. GFI offers post-hoc, generator-specific repair for frozen, black-box detectors, with three distinguishing properties:

  • More effective post-deployment Unlike NS-Net/LoRC, which improve generalization upfront, GFI targets specific blind spots after deployment. With 100 images, fine-tuning improves mean AUC from 0.78 to 0.82, while GFI reaches 0.93 (FPR@R95: 0.47 vs. 0.24), without updating the detector.
  • Real-image-only residual space Unlike NS-Net/LoRC, which use text-derived subspaces or supervised fake–real training, GFI learns its subspace once from unlabeled real photos and reuses it across generators, lowering fingerprint-learning costs and extending detector coverage.
  • Closed-form, extensible enrollment PCA and real-only calibration estimate fingerprints without gradient updates. As detailed in 1bgK (#2), just 50–100 generated images yield meaningful gains. Fingerprints can be added or removed independently, without rebuilding existing entries.

GFI is not a replacement for detectors like NS-Net/LoRC, but a complementary layer that inexpensively extends coverage by repairing their blind spots.

2. PCA Isolation vs. Full Prototypes and Nearest-Prototype Classifiers

Standalone fingerprint AUC (App. D.2) hides calibration and fusion effects. With fusion, a full-feature prototype (γ=0\gamma=0) is dominated by real-image variation, widening real-score swings (Table 2), increasing false alarms. Nearest-prototype classifiers share this flaw. PCA removes dominant real-image variation, leaving a quiet space where subtle generator traces stand out. Two findings validate this:

  • Generator signal lives in low-variance residual: Replacing a fake image's top-95% PCA components (semantic content) with real-image components barely changes fingerprint AUC (<0.001), whereas perturbing the remaining 5% residual destroys detection (App. D.1). Split-sample Crossnobis across different ViTs confirms that removing 95% of real-image variance still preserves 73–89% of generator separation:

    Encoder (K95K_{95})CLIP (146)DINO (98)AugReg (117)MAE (120)
    Naive/Crossnobis.783/.775.893/.887.766/.733.853/.842
  • PCA improves fused detection and rescues fakes missed by the base:

    DetectorNo PCA (γ=0\gamma=0)γ=.50\gamma=.50γ=.90\gamma=.90γ=.95\gamma=.95Missed-Fake Rescue (γ=0→.95\gamma=0 \rightarrow .95)
    DDA.931/.241.956/.161.932/.229.972/.127.566 →\rightarrow .907
    Effort.917/.248.943/.172.944/.162.914/.266.542 →\rightarrow .931
    SPAI.848/.471.873/.397.862/.481.900/.330.438 →\rightarrow .707
    Mean.899/.320.924/.243.912/.291.929/.241.513 →\rightarrow .841

    Sweeping γ\gamma shows PCA outperforms γ=0\gamma=0 across detectors for $\gamma \in [0.9,0.95]$ and on aggregate at γ=.95\gamma=.95, though the optimum varies by detector–generator pair. At γ=.95\gamma=.95, GFI rescues 84% of missed fakes versus 51% at γ=0\gamma=0, across 9/11 blind spots (the two exceptions are CycleGAN, where γ=0\gamma=0 was near ceiling).

    We also evaluate a nearest-prototype classifier (fake vs real centroid distance); it performs on par (≤.01\le.01AUC, FPR) with the full-feature mean prototype γ=0\gamma=0 and falls short of PCA-isolated GFI.

Together, these results show that PCA isolates generator signals that full-feature prototypes obscure, improving detection and recovering substantially more missed fakes.

3. Bank Scalability

Enrolling all 13 available AIGCDetectBench generators (N=100,γ=.95N=100, \gamma=.95, 5% real threshold):

Metric (DDA/Effort/SPAI)Empty (0)Targets (3–4)Full Bank (13)
Target AUC.827/.747/.762.972/.914/.900.939/.905/.875
8-Domain OOD FPR.083/.132/.106.111/.153/.164.102/.153/.133

Adding unrelated entries incurs a modest AUC drop with no extra OOD FPR growth compared to target-only enrollment, keeping detectors well above the empty bank.

4. Unknown/Inaccessible sources; Few-Shot FT

As detailed in 1bgK #5, GFI does not require generator identities for enrollment. We demonstrate this by clustering unlabeled images in residual CLIP space, where images with matching fingerprint directions naturally group together, allowing black-box enrollment of known and unknown generators. We also compare GFI and fine-tuning under the same 100-image budget across all 11 blind spots (Table 3). Few-shot fine-tuning improves mean AUC/FPR@R95 from 0.78/0.58 (Base) to 0.82/0.47 (FT), while GFI achieves 0.93/0.24. FT implementation and analysis are shared in 7Vjw #2.

Thank you for the clarifications and additional experiments. Although some concerns about methodological novelty and false positives under distribution shifts remain, I have increased my score by one point.

I also suggest improving the figure, as some illustrations appear AI-generated and detract from the paper’s overall presentation.

Thank you very much for taking the time to review our rebuttal, and raise your score—we really appreciate it.

Regarding the figure: thank you for pointing it out. We will replace the illustrations with a clean diagram.

On the remaining points regarding methodological novelty and false positives under distribution shifts, can you kindly share if there are specific aspects, comparisons, or cases you would like us to clarify or evaluate further before the discussion closes? We are happy to provide any additional analysis or discussion that would help fully resolve these concerns.

Thank you again for your constructive guidance.

Reviewer_7VjwOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Generator Fingerprint Isolation (GFI), a training-free method for repairing generator-specific blind spots of frozen AI-generated image detectors. GFI removes dominant principal components of real-image embeddings, constructs fingerprints from target-generator samples in the residual space, and fuses calibrated fingerprint scores with the original detector’s predictions. Experiments on AIGCDetectBench suggest that GFI improves detection performance on targeted generators while maintaining average performance on non-target generators.

Critical strengths
  1. The proposed GFI repairs known generator-specific blind spots without updating the detector. Such blind-spot repair can be a practical paradigm to compensate for the limited generalization of AIGI detectors.
Critical weaknesses
  1. The main experiments are conducted on AIGCDetectBench, where the generators are outdated. It remains unclear whether the proposed method can address the blind spots regarding more advanced generators, including open-source models such as the FLUX series and proprietary models such as GPT Image and Nano Banana series.
Questions for authors
  1. How is the fine-tuning (FT) baseline implemented? Why does FT yield little improvement for Effort and SPAI, as suggested by Table 3?
  2. Robustness to image post-processing: How does GFI affect detector robustness to common image transformations, such as JPEG compression, resizing, and Gaussian blur?
3 comments

We thank the reviewer for their feedback and address each concern below.

1. GFI + Contemporary Generators

We evaluated GFI with frozen Effort and SPAI on three modern open-source text-to-image models—FLUX.2-klein-4B, SD3.5-Large, and Qwen-Image—using 100 enrollment images and 300 test pairs per model for Effort, 200 for SPAI. DDA was excluded because its frozen baseline already reached ~0.99 AUC on all three. GPT Image and Nano Banana were not evaluated due to commercial API access constraints during the rebuttal period. Each fake was generated from the caption of a held-out FFHQ image, which served as its paired real test image.

AUC / FPR@R95 (Base → GFI)

DetectorFLUX.2-kleinSD3.5Qwen-Image
Effort.697/.677 → .938/.447.853/.450 → .982/.117.704/.663 → .930/.407
SPAI.705/.855 → .905/.475.968/.125 → .992/.020.860/.815 → .937/.400

GFI improves AUC and lowers FPR@R95 across all six detector–generator pairs.

2. Fine-tuning (FT) Analysis

In Table 3, FT used N=100N=100 fakes per target (pooled per detector) and an equal number of real images, disjoint from AIGCDetectBench. Training ran up to 5 epochs with early stopping, which selected epoch 1 across detectors due to rapid overfitting on small data:

  • DDA: Rank-8 LoRA on DINOv2 attention/MLP + head; AdamW, LR 10−410^{-4}, BCE loss.
  • Effort: Trainable SVD residuals of attention weights + 2-class head; AdamW, LR 10−410^{-4}, BCE loss.
  • SPAI: Frozen ViT; trainable SRS, SCA, and classifier; LR 10−410^{-4}, BCE + distillation (α=0.5\alpha=0.5).

Across all 11 blind spots (Table 3), mean AUC/FPR@R95 is 0.78/0.58 (Base), 0.82/0.47 (FT), and 0.93/0.24 (GFI). FT helps DDA, whereas Effort and SPAI gain little for different reasons—a data/optimization limit for Effort and an architectural limit for SPAI:

Why Effort gains little We measured how well each generator's mean fake–real shift in Effort's frozen feature space aligns with its learned classifier direction. For Effort's 3 blind spots the shift is nearly orthogonal to the classifier (mean cosine 0.114), so only a small part of it lands on the classifier axis (projected magnitude 1.65); for the 4 generators Effort already detects, alignment is higher (cosine 0.288, projection 5.32). The gap is orientation, not raw shift strength: e.g.: WFIR's shift is large (21.03) but nearly orthogonal (cosine 0.061), so almost none of it projects onto the classifier. Rotating Effort's residual weights onto these directions needs more data—with 4×4\times more, FT recovers:

DetectorSetupAUC (Base → FT)
Effort (N=400N=400)Adam 2×10−42\times10^{-4}, up to 5 ep.StyleGAN2: .69→.95; WFIR: .75→.998; ADM: .80→.928

Why SPAI struggles SPAI pretrains its ViT on real images only via self-supervised masked spectral learning and keeps that ViT frozen, while downstream modules (SRS, SCA, and classifier) score how poorly an image’s spectrum reconstructs. Because the ViT is frozen, FT can only reweight existing spectral cues; it cannot learn new features when a generator’s spectrum already resembles real images. Even at N=400N=400, downstream FT helps some targets while degrading others:

DetectorSetupAUC (Base → FT)
SPAI (N=400N=400)SRS/SCA/head, LR 2×10−42\times10^{-4}, up to 5 ep., ViT frozenADM: .81→.907; WFIR: .58→.673; CycleGAN: .87→.838; StyleGAN: .79→.733

Unfreezing SPAI’s ViT would disturb its real-image spectral model and require full retraining, whereas GFI repairs all three detectors at N=100N=100 without changing detector weights.

3. Robustness to post-processing

We evaluated Effort (Base → GFI) on the test fakes and paired reals from #1 under JPEG compression, 0.5×0.5\times bilinear resize/restoration, and Gaussian blur. Real-image calibration (RcalR_{cal}) was kept fixed, and the threshold was set per condition for 95% recall (SPAI was omitted due to its multi-patch runtime across 21 conditions):

ConditionFLUX.2-kleinSD3.5Qwen-Image
Clean.697/.677 → .938/.447.853/.450 → .982/.117.704/.663 → .930/.407
JPEG Q90.686/.727 → .905/.458.835/.493 → .957/.183.698/.660 → .919/.413
JPEG Q70.693/.710 → .898/.455.819/.443 → .952/.207.735/.557 → .901/.387
JPEG Q50.728/.753 → .892/.468.846/.447 → .949/.223.739/.583 → .890/.420
Resize 0.5.705/.683 → .877/.480.851/.440 → .965/.167.732/.643 → .902/.423
Blur σ=1\sigma=1.701/.687 → .884/.463.855/.447 → .969/.160.726/.663 → .892/.427
Blur σ=2\sigma=2.711/.707 → .871/.433.854/.427 → .953/.280.757/.600 → .891/.420

GFI improves over base across all conditions, with moderate attenuation under heavy compression and blur. Enrolling post-processed samples during fingerprint extraction is a natural extension.

Reviewer_7VjwOpenReview ↗

Followup questions and suggestions

Thank you for the responses. The supplemented results on contemporary generators have addressed my concern about the limitation in the tested generators.

Regarding the robustness to image processing, the performance degradation of GFI looks acceptable, but the comparison with the baseline is intriguing. Specifically, the baseline AUC is not significantly deteriorated by the processing in most cases, and may even increase (e.g., for JPEG Q50 on FLUX.2 and Qwen-Image). The comparison of GFI and the baseline seems to suggest that GFI is more sensitive to such image processing. Can you explain this phenomenon?

Besides, as minor suggestions, the formality of the manuscript should be improved. For example, Figure 2 lacks sufficient visual clarity, and the abstract should be a single paragraph.

We thank the reviewer for the follow-up question.

We agree that GFI's fingerprint channel is more sensitive to post-processing than the base detector. This partly reflects differences in training and enrollment: Effort's paper ($\S$4.1) explicitly states that its training uses several data augmentations, including Gaussian blur and image compression, whereas GFI's fingerprints and real-image calibration are based on clean images, with transformations introduced only at evaluation. Consequently, post-processing can alter the extracted fingerprint features, making them less aligned with the fingerprints enrolled from clean images, while the base detector has encountered similar perturbations during training.

GFI nevertheless provides complementary gains under the tested conditions, although its fingerprint contribution is more sensitive to processing. We will clarify this limitation and investigate transformation-aware enrollment and calibration in future work.

We will also improve Figure 2's visual clarity and present the abstract as a single paragraph.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.