acceptodds
Withdrawn from ICLR 2027

Forensic Twins: Self-Supervised Residual Learning for AI-Generated Image Forensics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Detectors of AI-generated images are typically trained using samples from all Generative AI architectures they must catch, and struggle as soon as a new architecture emerges. Recent approaches have explored self-supervised pre-training as an alternative solution, yet standard frameworks work against the forensic task, e.g., their augmentations overwrite the micro-statistics of image formation. This paper introduces Forensic Twins, a Self-Supervised Residual Learning (SSRL) framework whose pretext task suppresses macroscopic content availability. Each image is mapped through a frozen, off-the-shelf forensic residual extractor, from which two spatially disjoint crops are drawn. Sharing no pixel, the two views retain minimal semantic structure to align, leaving a redundancy-reduction objective with a predominant common signal: the stationary fingerprint of the image acquisition pipeline. Additionally, Forensic Twins is trained exclusively on real images; no AI-generated image is observed at any stage. Experiments show that Forensic Twins attributes AI generator sources with 56.61% accuracy, i.e., 6.13% above the previous state-of-the-art zero-shot method at lower latency. We also demonstrate that fitting a Gaussian Mixture Model (GMM) offline using only the real image embeddings extracted from Forensic Twins turns it into a state-of-the-art zero-shot detector, reaching 97.99% AUC across 27 unseen AI generators, including GANs, diffusion models and commercial systems. Code, weights and exact splits will be made publicly availableThe link will be provided upon acceptance..

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Media Forensics, Self-Supervised Learning, Residuals, AI-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_2JzZOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Forensic Twins, a self-supervised residual learning (SSRL) framework for AI-generated image detection and source clustering. A frozen forensic residual extractor suppresses high-level semantic content, and a lightweight ViT is trained with the Barlow Twins objective on two spatially disjoint crops, encouraging it to capture shared acquisition-related micro-statistical patterns. Pretrained on approximately 500,000 real images, the model achieves 97.99% AUC across 27 unseen generators using a one-class GMM fitted to its embeddings. With 3.0 million parameters and a reported inference time of 6.44 ms per image, it demonstrates strong generalization across GAN-based, diffusion-based, and commercial generators.

Critical strengths

The core contribution is achieving 97.99% AUC across 27 unseen generators after pretraining on approximately 500,000 real photos, without exposure to generated images. This suggests that real-image acquisition statistics provide strong cues for detecting AI-generated content under the evaluated settings.

Critical weaknesses
  1. The motivation and notation for residual extraction need clarification. Section 3.1 does not clearly explain the relationship between (I_{\text{res}}) and (X_{\text{res}}) or their correspondence to the processing steps in Figure 1. Please clarify these definitions and the rationale for residual extraction.
  2. The choice of (K=1) for the GMM is insufficiently justified. Given the diversity of real-image content and sources, please explain why a single Gaussian component is appropriate and provide a sensitivity analysis over (K).
  3. Potential overlap between model selection and evaluation requires clarification. Both the ablations and main evaluation use the same 27 generators. Were their results used to select architectures or hyperparameters? If so, evaluation on separate generators would strengthen the zero-shot generalization claims.
  4. The contribution of FRE is not fully isolated. In Table 3, subsequent architecture and hyperparameter studies retain FRE, but the final configuration is not evaluated without it. This comparison would help distinguish the contribution of FRE from that of the remaining framework.
Questions for authors

Could the authors report seed-to-seed variability or confidence intervals for the main results? Since the evaluation covers diverse generators, this would help clarify the stability of the reported performance.

Reviewer_QPSpOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose to alleviate the limitations of a prior work on identifying forensic microstructures for image forensics, FSD (CVPR 2025) where the residual images inherently retain structural semantic edges that can be leveraged by the encoder to create an optimization shortcut. They alleviate this by introducing spatially disjoint views for an image within the training pipeline that uses a self-supervised objective function to maximize the invariant factors between two views of an image. The learnt representations and residuals are expected to be invariant to semantics and aid two forensics tasks: synthetic image detection and image source attribution.

Critical strengths

The paper tackles an important and a highly-pertinent problem in image forensics, with the intent to develop a representation that not only detects real from synthetic videos, but also enables the attribution of the source of an image (generated; generator type, camera pipeline) etc. Such generalizable representations are quite difficult, especially considering that an image / video typically encountered over the internet would go through various transformations that the representations must be robust to. The fact that Forensic Twins, by design, has the potential to enable representations that can answer multiple forensics questions is very impactful in my opinion.

Critical weaknesses

The novelty over the CVPR 2025 paper seems worth a consideration: the whole proposal this method rests on is that the CVPR'25 paper (L213-215) produces residual images which inherently retain structural semantic edges. Please consider backing this with: a. showing that FSD is actually taking a semantic shortcut (example: one could cluster FSD's representations by ImageNet classes to show consistency between classes and residuals; and this clustering should reduce drastically with the new approach) and b. showing a reduction in mutual information between feature representations for various classes of semantically-consistent images (e.g., such as ViT feature representations of ImageNet classes) when compared to the proposed approach vs. that with FSD.

The overall approach is aimed at isolating the camera-pipeline micro-signatures vs the ones left behind by a synthetic image generator. This needs to visualized to motivate better: do these two signals indeed differ significantly? the micro-statistical noise is somewhat ill-defined / deserves a visual aid. Moreover, the camera-processing micro-statistical noise that the authors intend to isolate seems susceptible to image / video compression. Please evaluate robustness along compression, re-encoding, and noise (perturbations which are very common and can impact the micro-statistics) at quality levels typically encountered on the internet (such as crf40 for videos).

Questions for authors
  1. Please verify that the image interpolation footprint of non-invasive image transformations with camera / forensic microstructure that is intended isolated by the method.

  2. Echoing from the above section here for completeness: The novelty over the CVPR 2025 paper seems worth a consideration: the whole proposal this method rests on is that the CVPR'25 paper (L213-215) produces residual images which inherently retain structural semantic edges. Please consider backing this with: a. showing that FSD is actually taking a semantic shortcut (example: one could cluster FSD's representations by ImageNet classes to show consistency between classes and residuals; and this clustering should reduce drastically with the new approach) and b. showing a reduction in mutual information between feature representations for various classes of semantically-consistent images (e.g., such as ViT feature representations of ImageNet classes) when compared to the proposed approach vs. that with FSD.

  3. The overall approach is aimed at isolating the camera-pipeline micro-signatures vs the ones left behind by a synthetic image generator. This needs to visualized to motivate better: do these two signals indeed differ significantly? the micro-statistical noise is somewhat ill-defined / deserves a visual aid.

  4. Continuing point 3: the camera-processing micro-statistical noise that the authors intend to isolate seems susceptible to image / video compression. Please evaluate robustness along compression, re-encoding, and noise (perturbations which are very common and can impact the micro-statistics) at quality levels typically encountered on the internet (such as crf40 for videos).

  5. GMM estimation: ViT [CLS] token vectors are high-dimensional. GMMs struggle with covariance estimation in high dimensions, leading to numerical instability or over-regularization. How do the authors alleviate this?

  6. Results visualization in Table 1: please highlight the best in bold. From table 1, the advantage of forensic twins over FSD is not apparent.

  7. Minor suggestion: The intro section reads a bit like tokenized text, with a large list of literature without providing the overall motivation / key insight that drives the method. Please consider editing for readability.

  8. Curious to know if this approach generalizes beyond images? How about applying it per-frame to videos?

Reviewer_bKg3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Forensic Twins, a self-supervised residual learning framework for AI-generated image detection and source clustering. The method combines a pretrained forensic residual extractor, spatially disjoint crops, and a Barlow Twins objective to learn image-formation fingerprints exclusively from real images. By suppressing semantic information, the framework aims to capture low-level forensic characteristics that generalize to unseen generative architectures. The learned representations are evaluated on two downstream tasks: zero-shot AI-generated image detection using a one-class Gaussian Mixture Model (GMM) and unsupervised image-source clustering using k-Means. The method achieves an average AUC of 97.99% across 27 generative architectures and demonstrates competitive clustering performance with relatively low computational overhead. The experimental results suggest that learning forensic representations from real images alone is a promising alternative to conventional supervised detection approaches.

Critical strengths
  1. Meaningful and well-motivated methodological innovation. The paper presents an interesting perspective on self-supervised learning for image forensics. Rather than relying on conventional semantic representations, the proposed framework encourages the learning of image-formation fingerprints through forensic residual extraction and spatially disjoint views. This task-specific design is well motivated and represents a meaningful contribution to forensic representation learning. I particularly appreciate the idea of adapting the self-supervised learning process to preserve forensic evidence while reducing reliance on semantic shortcuts.

  2. Competitive generalization performance with high computational efficiency. The ability to achieve competitive detection performance across 27 generative architectures while training exclusively on real images is a notable strength. This property is particularly valuable given the rapid evolution of generative models. Furthermore, the method achieves 97.99% average AUC with only 3.0M parameters and 6.44 ms inference latency. Compared with FSD, it provides approximately 375 times faster inference while maintaining comparable detection performance. The extensive comparisons and ablation studies further support the practical relevance of the proposed approach.

Critical weaknesses
  1. Limited robustness to common image post-processing operations. The substantial performance degradation under JPEG re-encoding raises concerns about the practical applicability of the proposed method. As reported in Appendix A, the detection AUC decreases from 97.99% on native images to 71.93% at JPEG quality 95 and 59.92% at quality 50. Although the authors acknowledge this limitation and investigate few-shot calibration, the original real-only detection framework remains sensitive to common compression operations. Since image re-encoding frequently occurs in real-world applications, this limitation is particularly concerning. Moreover, the current evaluation does not provide a direct comparison with competing methods under identical JPEG compression settings. Such a comparison is important for determining whether this is a general limitation of forensic detectors or a more pronounced weakness of the proposed approach.

  2. Insufficient isolation of the proposed learning strategy's contribution and its learned representations. The framework combines an existing forensic residual extractor with the established Barlow Twins objective. Although Table 3 provides useful ablation studies, a substantial portion of the performance improvement appears to come from the pretrained residual extractor. Comparisons with alternative self-supervised objectives under identical residual inputs and encoder architectures are still missing. These experiments would help establish the independent contribution of the proposed learning strategy. Furthermore, although the semantic-versus-source analysis provides encouraging evidence, the current experiments do not fully rule out the influence of content-dependent texture statistics or dataset-specific biases. More controlled evaluations would strengthen the claim that the learned representations primarily capture image-formation fingerprints rather than semantic correlations.

  3. Minor presentation concerns. The paper is generally well organized and readable, but some expressions require additional proofreading. For example, “Both crops sizes” (Section 3.2) and “these elaborated residual extractors” (Section 3.1) could be improved. In addition, Tables 4 and 7–9 contain densely arranged results and relatively small font sizes, while some labels in Figure 1 are difficult to read. Improving the language consistency and visual presentation would enhance the manuscript’s overall clarity. These are minor issues and are not central to my recommendation.

Questions for authors

Q1. Contribution of the self-supervised learning strategy. Can the authors provide a controlled comparison using alternative self-supervised learning objectives under the same forensic residual extractor, encoder architecture, and training settings? In particular, how much of the improvement can be attributed specifically to the spatially disjoint view construction and the Barlow Twins objective, rather than the pretrained residual extractor?

Q2. Robustness under JPEG re-encoding. Appendix A demonstrates substantial performance degradation under JPEG re-encoding. How does this degradation compare with competing methods under identical compression settings? Can the authors provide additional evidence regarding whether the proposed method maintains its relative advantages after common post-processing operations? Reporting detection performance at practically relevant operating points, such as TPR at low FPR, would also help clarify its practical reliability.

Q3. Semantic–forensic disentanglement. Can the authors provide additional evidence demonstrating that the learned representations primarily capture image-formation characteristics rather than content-dependent texture statistics or dataset-specific biases? For example, experiments controlling semantic content across different image sources would help strengthen this claim and clarify the nature of the learned forensic representations.

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

Related papers

Discussion (0)

Sign in to comment.