acceptodds
Withdrawn from ICLR 2027

Prototype-Guided Tangent–Boundary Perturbation for AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Detectors that synthesize features during training have improved AI-generated image detection. However, our visualization reveals that synthesized features reshape rather than enrich the original fake-feature distribution, causing the detector to learn from the synthesized features while losing its understanding of the original distribution. To introduce additional training variations while retaining supervision on the original features, we propose Prototype-Guided Tangent-Boundary Perturbation (PTBP), a feature-level perturbation method. First, its orthogonal subspace decomposition uses class prototypes and within-class feature variances to decompose sampled Gaussian noise into components orthogonal and parallel to the discriminative direction. This decomposition allows PTBP to control what each perturbation component preserves. Then, controlled feature perturbation uses the orthogonal component to construct tangent views that preserve the predictions of the corresponding original features. It further adds a class-aware and margin-calibrated parallel component to construct boundary views, which are supervised with the original binary labels after class-aware shifts. The two views are trained using tangent consistency and boundary supervision, respectively. Both objectives are jointly optimized to encourage the detector to learn from additional training variations while retaining supervision on the original features. With only one training epoch, PTBP achieves the best performance in both cross-generator and compression-perturbed scenarios.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Feature-level perturbation, Orthogonal subspace decomposition, AI-Generated image detection

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

This paper targets cross generator generalization for AI-generated image detection. Observing that SimLBR generates synthetic features which reshape rather than enrich original fake feature distributions, the paper proposes PTBP, a feature-level augmentation method working on frozen DINOv3 features. It decomposes Gaussian noise via orthogonal subspace decomposition into components parallel and orthogonal to a global discriminative direction, constructing tangent views for prediction consistency and boundary views for hard sample training.

Critical strengths
  1. The motivation is clear. The paper points out a critical flaw of SimLBR, i.e. synthesized latent mixed features reshape the original feature manifold and dilute supervision from real/fake original samples.
  2. The design preserves supervision on original features; perturbations are auxiliary views instead of replacing training samples. Inference introduces zero overhead. Training is efficient.
  3. Comprehensive experiments are provided, including cross generator evaluation, compression robustness test, component ablation and t-SNE visualizations to support the effectiveness.
Critical weaknesses
  1. The discriminative direction is globally computed from class level real/fake prototypes and shared for all samples. DINOv3 features lie on highly nonlinear manifold; local decision normals vary per sample. Mathematically global orthogonal noise components cannot strictly guarantee label preserving tangent perturbation.
  2. It is feature space augmentation, which is an incremental improvement within an existing paradigm.
  3. PTBP only perturbs features in training phase. It does not handle false positive risks caused by normal responses from real images during inference. Score drift under severe test time distribution shift remains unsolved.
Questions for authors
  1. The discriminative direction is globally computed from class level real/fake prototypes and shared for all samples. DINOv3 features lie on highly nonlinear manifold; local decision normals vary per sample. Mathematically global orthogonal noise components cannot strictly guarantee label preserving tangent perturbation.
  2. It is feature space augmentation, which is an incremental improvement within an existing paradigm.
  3. PTBP only perturbs features in training phase. It does not handle false positive risks caused by normal responses from real images during inference. Score drift under severe test time distribution shift remains unsolved.
Reviewer_AZ4YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Prototype-Guided Tangent–Boundary Perturbation (PTBP) for generalizable AI-generated image detection. Using features from a frozen CLIP encoder, the method defines a discriminative direction from class prototypes and within-class variances. Gaussian noise is decomposed into orthogonal and parallel components to construct tangent and boundary views. The feature fusion module and classifier are trained jointly using original classification, prototype regularization, tangent consistency, and boundary supervision. Experiments evaluate cross-generator generalization and JPEG compression robustness.

Critical strengths
  • The paper addresses an important practical problem: improving generalization using available training data without collecting images from additional generators.
  • The complete method reports a substantial improvement over its own baseline and higher mean accuracy across four JPEG evaluation settings.
  • The perturbation procedure is disabled during inference, so prediction does not require constructing additional views.
Critical weaknesses
  • The central motivation and distribution-preservation argument are insufficiently supported. The claim that synthesized features cause a detector to lose its understanding of the original distribution is inferred primarily from centroid displacement and reduced within-class variance. These observations do not directly establish forgetting or a detrimental causal effect on generalization. A crucial control would be to retain classification supervision on the original fake features in SimLBR, thereby distinguishing the benefit of preserving original supervision from that of PTBP’s perturbation design. Furthermore, PTBP’s own prototype regularizer encourages within-class contraction. The paper should distinguish retaining supervision on original samples from preserving their distributional structure, and explain why the contraction encouraged by its regularizer is compatible with its motivation.
  • The geometric interpretation of tangent and boundary perturbations lacks direct validation. The prototype-derived direction is not necessarily aligned with the actual classifier’s boundary normal, and class-specific scaling further changes the perturbation geometry. The authors acknowledge that the constructed tangent hyperplane need not align with the classifier boundary. Consequently, orthogonality to the prototype direction does not ensure prediction stability, while movement along that direction does not necessarily move a sample toward the current decision boundary. Likewise, normalizing distances by the batch median and clipping the resulting coefficients does not constrain the resulting classification margins. The paper should directly measure prediction changes under tangent perturbations and margin changes under boundary perturbations, with magnitude-matched random perturbations as controls. The current evidence does not establish whether the two views consistently serve their intended roles or when their supervision assumptions remain appropriate.
  • Since training fake images come exclusively from ProGAN, the prototype difference may capture source-generator-specific cues, semantic differences, or dataset biases rather than transferable forensic information. Representing each class with a single prototype also imposes a strong simplification on potentially multimodal feature distributions. Although the compression experiments vary the number of semantic categories, the paper does not directly analyze how the prototype-derived direction changes across these settings or whether it remains reliable under a different training generator.
  • The ablations and controls do not isolate the contributions of the proposed designs. Table 6 compares the baseline, unstructured Gaussian noise, orthogonal decomposition, and the complete method. However, it does not separately evaluate prototype regularization, tangent consistency, and boundary supervision. The training objective and view construction of the “decomposition-only” variant are also insufficiently specified. Moreover, it is unclear whether the Gaussian-noise baseline matches PTBP in perturbation magnitude, class-dependent scaling, number of views, and auxiliary objectives. The observed difference therefore cannot be attributed specifically to directional decomposition. Individual loss ablations, separate evaluation of the two views, and comparisons that change only the perturbation direction or magnitude-calibration rule are needed to establish the incremental value of the complete design.
  • The compression robustness gains lack an architecture-matched control. Tables 2–5 compare methods with different architectures and pretrained representations, with some comparison results taken from prior work. The paper does not directly compare the same frozen CLIP encoder, fusion module, and classifier with and without PTBP under the identical JPEG training protocol. The cross-generator ablation cannot substitute for this comparison. The reported results demonstrate that the complete system performs well under the evaluated compression settings, but do not sufficiently isolate PTBP’s contribution. In addition, the compression evaluation is dominated by GAN-based sources, leaving its robustness across other generation mechanisms insufficiently tested.
  • Important implementation details and training-stability evidence are missing. The paper does not adequately specify the CLIP feature extraction layers, token representations, fusion architecture, prototype initialization, or several important hyperparameters, including perturbation strength, perturbation probability, EMA momentum, temperature, and the upper bound on the margin factor. These choices directly affect the proposed augmentation and limit reproducibility. The paper also lacks repeated training runs, sensitivity analyses for the principal hyperparameters, and learning curves. The standard deviation in Table 1 measures variation across generators rather than training seeds. Reporting only one epoch does not establish whether the baseline would catch up with further training or whether PTBP would deteriorate as training continues. These omissions limit confidence in the stability of the reported improvements.
  • The evaluation scope and metrics limit the breadth of the conclusions. The main generalization results rely on a fixed ProGAN training protocol, without validation under alternative training generators or sufficiently independent data sources. The paper primarily reports accuracy, without AUC, AP, or real-image false-positive rates, making it difficult to distinguish improvements in ranking performance from threshold-specific effects. The t-SNE visualizations provide qualitative illustrations but cannot establish that the model learns transferable forensic evidence. The experiments support improvements under the evaluated protocols; broader claims about generalization and reliability require more direct evidence.
Questions for authors
  • Under matched losses, class-dependent scaling, and perturbation magnitudes, how does PTBP compare with random-direction perturbations and perturbations constructed from the actual classifier boundary?
  • How do tangent perturbations affect predictions, and how do boundary perturbations change the signed classification margin with respect to the original label?
  • How do SimLBR with retained original-fake supervision and PTBP with individual auxiliary losses removed perform?
  • What are the results of the same CLIP-based detector without PTBP under all four JPEG evaluation settings?
Reviewer_iDvgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PTBP, a feature-space perturbation method for improving cross-generator generalization in AI-generated image detection. It decomposes Gaussian noise using real/fake prototypes to construct tangent and boundary views, optimized through consistency and classification losses. Experiments using a frozen CLIP backbone report 94.76% mean accuracy across 16 generators and improved robustness under JPEG compression.

Critical strengths

S1. A lightweight training-time method. PTBP acts only on features during training and adds no cost at inference. It is evaluated under two established protocols covering 16 generators and four JPEG compression settings.

S2. Competitive mean accuracy. PTBP achieves the highest mean accuracy in Table 1 (94.76%) and in all four compression settings, exceeding DDOC by 3.4 to 8.0 points. These are single-run results, however, and PTBP does not lead on the standard-deviation or reliability metrics.

Critical weaknesses

W1. The gains are not convincingly attributed to the proposed perturbations. Fig. 1(d) indicates that the perturbations amount to only a few percent of the feature RMS, and each auxiliary loss is weighted by 0.01. Unstructured Gaussian noise leaves the baseline essentially unchanged (84.96% vs. 84.89%). Because the views are applied immediately before a linear classifier, small additive perturbations of this kind act mainly as weight regularization (Bishop, 1995), so it is unclear how they yield the 4.05-point gain of the full model over the OSD-only variant. The prototype regularizer is a center-loss-style term that may explain much of the jump from the baseline to the OSD-only variant, but it is never ablated on its own. It also explicitly reduces within-class variance, which sits uneasily with the motivation in Fig. 1(c). All results come from a single one-epoch run, while individual generators shift by 20 to 30 points across ablation variants, as on GauGAN (96.16, 65.46, then 84.37), so variance across seeds is needed.

W2. The geometric construction does not guarantee the properties that motivate it. The authors acknowledge that the tangent hyperplane need not align with the classifier normal. In addition, u is defined in the space whitened by the shared matrix D, whereas orthogonality is enforced in the metric scaled by the class-specific D_y. Tangent views can therefore move along the discriminative direction in the very space where it is defined. The margin factor ρ_i is normalized by the median distance rather than by each sample’s own distance to the boundary. Boundary views are thus not guaranteed to stay on the correct side of the decision boundary, in which case supervising them with the original label introduces label noise. None of these properties is measured.

W3. Evaluation and reporting. The internal baseline reaches 84.89%, about 7.6 points below RINE in the same table, which also builds on intermediate CLIP features. The 9.87-point improvement is therefore measured against a weak reference; relative to RINE, the gain is 2.27 points. The source of the Table 1 baseline numbers and the ProGAN training subset are not stated. The compression experiments include no ablation and do not describe how PTBP uses the paired and unpaired training data. Several transcribed baseline means do not match their row averages, such as ADD in Table 3 and FreqNet in Table 4. The perturbation strength λ, the perturbation probability p, the clipping bound ρ_max, the temperature T, the EMA coefficient m, and ε are not reported.

W4. Limited novelty and weak qualitative evidence. Class-conditional feature-space augmentation (ISDA; Wang et al., 2019), center-loss compactness (Wen et al., 2016), and consistency regularization are established techniques that are neither discussed nor compared. The t-SNE maps in Figs. 3 and 4 show nearly identical curve-shaped structures across different generators. This pattern is more suggestive of features concentrated on a low-dimensional manifold than of improved separation, and quantitative measures would be more informative.

Questions for authors

Q1. Given that the perturbations appear to be only a few percent of the feature RMS and each auxiliary loss is weighted by 0.01, how do the tangent and boundary views account for the 4.05-point gain of the full model over the OSD-only variant? It would help to report the average change in the logit margin that these views induce. Results averaged over several seeds would also clarify whether the differences in Table 6 exceed run-to-run variation. This is the question most likely to change my assessment.

Q2. What exactly does the OSD-only variant optimize, since orthogonal subspace decomposition by itself only produces noise components? Please report a variant trained with the prototype regularizer alone, as well as variants that remove the tangent consistency loss and the boundary supervision loss one at a time. I would also ask the authors to explain how a regularizer that compacts each class is consistent with the motivation that reduced within-class variance is harmful.

Q3. How well does the prototype direction align with the classifier normal during training? Please report how often tangent views change the prediction and how much they shift the logit margin. For boundary views, please report the fraction that cross the classifier’s decision boundary. A comparison with norm-matched perturbations defined directly along the classifier normal would show whether the prototype geometry is what matters.

Q4. Could the authors report the values of λ, p, ρ_max, T, m, and ε, describe how they were selected, and specify the ProGAN training subset? Please also state whether the Table 1 baselines were trained on the same data or transcribed from prior work. Since the internal baseline is about 7.6 points below RINE, which also uses intermediate CLIP features, an explanation of this gap would help in judging the size of the improvement.

Q5. Could the authors provide the component ablation under the compression protocol and explain how PTBP uses the paired and unpaired portions of the training data? Comparisons with CLIP-based detectors trained under the same compression protocol would help separate the effect of PTBP from that of the backbone.

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

Related papers

Discussion (0)

Sign in to comment.