acceptodds
Under review as a conference paper at ICLR 2027

Continuous Syndrome Forensics for Proactive Image Manipulation Localization and Reconstruction

Abstract

The proliferation of generative AI motivates proactive image forensics capable of accurately localizing manipulations and reconstructing visually plausible semantic content. However, existing proactive defenses primarily rely on empirical deep learning, lacking principled algebraic foundations and remaining fragile under global distortions. While classical coding theories offer rigorous paradigms, their direct adaptation to natural images is non-trivial: finite-field error-correction codes are not directly applicable to continuous-valued pixels without discretization, and tools like classical compressive sensing commonly require strict global sparsity that fails under large-area edits, alongside dense linear projections that degrade visual quality. To bridge this gap, we propose Continuous Syndrome Forensics (CSF), which extends syndrome-based verification to the continuous spatial domain. By imposing block-wise algebraic invariants, CSF introduces an algebraically structured formulation for tamper localization without global sparse-recovery optimization. Although it is effective for small edits, directly applying CSF to natural images suffers from host self-interference and remains underdetermined under large-area manipulations. To address these challenges, we further introduce SyndromeMark, a network-augmented framework. SyndromeMark leverages the continuous syndrome residual proposed in CSF as an algebraically anchored spatial prior to guide both localization and reconstruction. Specifically, this prior enables a localization network to separate tampering from host-texture leakage, while guiding a reconstruction network to synthesize contextually plausible content consistent with the algebraic constraint. Extensive evaluations show that SyndromeMark improves robustness over the evaluated proactive baselines under severe post-processing distortions, supports both localization and semantic reconstruction, and remains resilient in the adaptive surrogate setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AIGC forensics, image manipulation, tampering localization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TEQkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper recasts proactive tamper localization as continuous syndrome verification. A block-diagonal semi-orthogonal matrix H, shared across 2x2 RGB patches, binds the watermarked image to the linear invariant HI_wm = G(m), where G(m) is derived from a 128-bit message. The back-projected residual f_syn = H^T (HI_atk - G(m)) is supplied to a U-Net localizer as extra input channels, and a second U-Net inpaints the masked region, supervised by MSE+LPIPS plus a syndrome regularizer. Training uses 5,000 Stable Diffusion images and testing is zero-shot on PASCAL VOC (splicing, copy-move) and CelebA (deepfake) under blur, noise, JPEG and patch removal, against three passive and four proactive baselines. The appendix provides four linear-algebra results, implementation details and extended tables.

Critical strengths
  1. The idea is simple, reference-free at verification time (no original image is needed), and easy to attach to an existing localizer. If the invariance held, this recipe would be broadly reusable.

  2. The zero-shot protocol — training only on synthetic images and testing on natural datasets without fine-tuning — is a genuinely strong generalization result, and reconstruction PSNR improves consistently across distortions and across both datasets.

Critical weaknesses
  1. The algebraic invariant is not established, and the evidence is internally inconsistent. Because the local block maps 12 scalars to 8 syndrome values, the constraint fixes a large majority of the image dimensions while G(m) depends only on a 128-bit message; the admissible set is therefore a single affine subspace, and an unseen host cannot lie near it while retaining the reported fidelity. Relatedly, the near-exact residual reported in Table 6 cannot coexist with the low syndrome SNR of Table 2, which is precisely the mismatch-dominated regime that the paper itself calls "Curse 1". If the embedding mismatch dominates, then f_syn is essentially a fixed random projection of the attacked image plus a constant : a high-pass channel and the ablation of Figure 4 cannot distinguish the claimed algebraic contribution from that channel. A control using a fixed high-pass or noise residual, and a control using an untrained random H, are both missing. Theorems 1-3 assume the constraint holds exactly and therefore do not cover the deployed system, as the authors themselves acknowledge.

  2. The empirical evidence does not support the paper's claims, and parts of it are self-contradictory.

(a) Head-to-head, the method does not win against the baselines the authors themselves list. On CelebA, RecoverMark is clearly better under Gaussian noise, JPEG and no distortion (F1 0.7870 vs 0.4350 under JPEG; 0.9746 vs 0.5057 with no distortion). On PASCAL VOC, the passive RelayFormer wins four of five settings (0.9554 vs 0.8566 under JPEG), and Table 3 of the paper itself already places a passive baseline above the proposed method under the same JPEG setting. RecoverMark is excluded from VOC on the grounds that it is face-specific, so the method only wins on the dataset where its strongest competitor is absent — and the dismissal of RecoverMark as depending on an unperturbed background is contradicted by precisely those noise and JPEG results.

(b) The motivating premise is asserted rather than measured, and the baseline set is too narrow to support a state-of-the-art claim. Recent IML methods such as Mesorch, PNF-Net, TruFor and ObjectFormer are trained and evaluated explicitly under compression, blur and noise, and degrade gracefully at the distortion levels used here, so the claim that passive forensics collapses under global post-processing is not established. The paper also never states whether the passive baselines were retrained with the same distortion augmentation and training budget as the proposed model, although the comparison turns on exactly that. The related work omits this robust-IML line and the closest proactive competitors (SepMark; LIDMark), which makes the claimed novelty over existing proactive forensics hard to assess.

(c) Three of the four stated system goals are effectively unverified. Invisibility is listed as a goal, but no PSNR/SSIM/LPIPS comparison of the protected image against any baseline is reported. Ownership verification is never compared with the extraction robustness reported by other watermarking methods: no bit-error rate, no payload accounting, and no wrong-key or unwarranted-claim test. Reconstruction is supported by PSNR alone, and the perceptual column is not on a consistent scale — the same Imuge+ is reported with an LPIPS an order of magnitude smaller on CelebA than on VOC at comparable PSNR, which makes the reconstruction comparison uninterpretable. The reconstruction network also never receives f_syn, so the claim that reconstruction is algebraically guided rests entirely on an unablated loss term.

Questions for authors
  1. On which data distribution was Table 6 computed? On the test sets, please report the relative magnitudes of the embedding mismatch, the projected tampering and the projected noise, and explain how an unseen host is mapped into the set defined by H*I = G(m) while keeping the reported fidelity.

  2. Is the message m a random identifier, a fixed payload, or optimized per image? What is the effective payload in bits, and how does it compare with the watermark capacity of the baselines?

  3. Please give the exact formula for the "spatial SNR" in Table 2: what are the numerator and denominator, is it computed on the pixel grid (f_syn ∈ R^n) or the syndrome grid (s ∈ R^d), and does it use the ground-truth mask? Please also state which λ_syn was used for Tables 1, 3, 4 and 5. Is the mask thresholded at 0.5? Are PSNR/LPIPS computed over the whole image or only over the tampered region, and why is Imuge+'s LPIPS an order of magnitude smaller on CelebA than on VOC at the same PSNR?

  4. Given AUC near 0.99 but F1 near 0.44, please report best-threshold F1, IoU and PR curves, along with the false-positive rate on clean but distorted images.

  5. Please add two controls for the localization network: one with a fixed high-pass/noise-residual channel instead of f_syn, and one with an untrained random H. Were the passive baselines retrained with the same distortion augmentation and training budget as the "without prior" variant?

  6. Which loss weights were used in the main tables, and does the syndrome term in the reconstruction loss matter at all? The reconstruction network never receives f_syn, so what makes reconstruction "syndrome-guided"?

  7. Please report watermark extraction bit-error rate against SepMark, EditGuard, LIDMark, OmniGuard and RecoverMark, together with wrong-key and unwatermarked-image false-claim rates, and report localization F1 under the adaptive attack rather than residual energy alone.

4 comments

Thanks for the detailed response. Before deciding the final score, I hope the author will briefly clarify the following three questions.

First, the paper's claim is about the algebraic invariant, but the evidence isolates only a small effect: 0.5478 AP against 0.5161 for a plain high-pass channel, with absolute AP around 0.55. What makes the system robust to distortions is the localizer and its training distribution, and that is not separated from the invariant.

Second, the paper still contradicts itself. The abstract and conclusion assert superiority while §6.2 concedes losses to RecoverMark and RelayFormer, and the operating point is 26.98 dB PSNR against roughly 37–44 dB for comparable methods, with no fidelity comparison.

Third, the comparison set remains selective: SepMark matches the payload and is the closest paradigm ancestor yet is neither cited nor compared, LIDMark appears only in the payload column, and Mesorch, PNF-Net, TruFor and ObjectFormer are cited without numbers. The stated motivation is also contradicted by the paper's own related work and Table 5.

AuthorsOpenReview ↗

Response to Reviewer TEQk (1/2)

Dear Reviewer TEQk,

Thank you for your follow-up comments. Below we clarify each of the three points and note the textual alignments made in our revised manuscript.

1. Misunderstanding Regarding the Algebraic Invariant and Empirical Isolation

There is a misunderstanding about what our algebraic invariant is and how its effect should be evaluated. Our algebraic invariant is the linear subspace constraint HIwm=G(m)HI_{wm} = G(m) embedded via a secret, block-diagonal semi-orthogonal matrix HH (HH⊤=IdHH^\top = I_d). By design, energy-bounded global distortions (enoisee_{noise}) are non-amplified under HH, whereas localized manipulations (etampere_{tamper}) violate the constraint and produce block-local syndrome energy spikes (s=HIatk−G(m)≈Hetampers = HI_{atk} - G(m) \approx He_{tamper}) that indicate tampered regions. This algebraic mechanism is fundamentally distinct from, and cannot be reduced to, a single pixel-level AP score.

Rather than relying on the localizer or its training distribution, the independent contribution of the algebraic invariant is established across multiple quantitative and qualitative experiments in the manuscript:

  • Table 9 (Zero neural localizer, zero LNet training distribution): Evaluating the raw syndrome residual ss without any neural network shows an embedding mismatch of only ∼10−6\sim 10^{-6}, while tampered blocks exhibit 13.55×13.55\times to 17.26×17.26\times higher mean residual energy than untampered blocks, achieving raw block-level ROC-AUCs of 0.7234 (PASCAL VOC), 0.7613 (Stable Diffusion), and 0.9328 (CelebA) across distortions.
  • Table 4 (Direct causal scaling with λsyn\lambda_{syn}): Increasing only the Phase 1 algebraic constraint weight λsyn\lambda_{syn} from 11 to 1000010000 boosts the spatial syndrome SNR in fsynf_{syn} from -1.91 dB (under blur) (and 5.18 dB under noise) up to 17.71 dB (Splicing) and 19.27 dB (Copy-Move), directly confirming that the forensic localization cue originates from enforcing HIwm=G(m)HI_{wm}=G(m).
  • Figure 8(a) (Pure linear CSF without LNet): Pure linear algebraic verification alone (with no neural localizer) already achieves ~0.95 F1 at a 10% tamper ratio.
  • Figure 5 (fsynf_{syn}-only input without RGB): Feeding LNet only fsynf_{syn} without the RGB image IatkI_{atk} yields ~0.85 F1 at lower tamper ratios, proving fsynf_{syn} carries standalone tampering cues rather than relying on RGB artifacts.
  • Table 2, Figure 4, and Figure 7 (Strictly controlled ablations under the identical localizer and training distribution): With the U-Net architecture, 5,000-image training set, and distortion augmentations held completely fixed, removing fsynf_{syn} (M^w/o\hat{M}_{w/o}) causes severe false positives under noise and JPEG compression (Figures 4 and 7). In Table 2, Syndrome Prior outperforms RGB-only, High-pass, and untrained Random Projection, cutting PASCAL VOC noise FPR from 2.83% (RGB-only) and 1.00% (High-pass) down to 0.17%, and reducing CelebA distortion-only FPR to 0.00% across all conditions.
  • Table 3 and Table 11 (Reconstruction constraint and key binding): Enforcing the syndrome term in Equation 8 reduces output constraint error MSEsyn\text{MSE}_{syn} by 32.01% (VOC) and 19.85% (SD) over standard inpainting (λsyn=0\lambda_{syn}=0, Table 3), and the true invariant residual remains orders of magnitude above the clean baseline under adaptive surrogate PGD attacks (1.698e-3 vs. 3.000e-6, Table 11).
AuthorsOpenReview ↗

Response to Reviewer TEQk (2/2)

2. Claim Consistency and Protected-Image Fidelity Comparison (Table 10)

  • Alignment of Abstract and Conclusion: We thank the reviewer for catching the legacy phrasing in Conclusion ("outperforming state-of-the-art baselines"). We have revised the manuscript so that the Abstract, Introduction, Section 6.2, and Conclusion consistently scope our claims over evaluated general proactive baselines.
  • Fidelity Comparison: SyndromeMark is a proactive method whose embedded redundancy trades off some visual fidelity for active verification and content recovery. Comparing protected-image PSNR against passive detectors, which do not alter images and offer neither ownership verification nor reconstruction, is not like-for-like. Meanwhile, 37 to 44 dB figures reported elsewhere correspond to different setups (e.g., 1024×10241024\times 1024 native resolution, fragile localization-only templates, or face-specific background offloading). Under our unified 256×256256\times 256 proactive evaluation, EditGuard itself measures 30.45–32.02 dB PSNR. In the latest revision (Appendix D.2, Table 10), when matched to EditGuard's fidelity regime (λsyn=100\lambda_{syn}=100: 31.65 dB on VOC, 32.70 dB on CelebA), SyndromeMark surpasses EditGuard in both perceptual quality (LPIPS 0.1469–0.1822 vs. 0.2273–0.3761; SSIM 0.8931–0.8988 vs. 0.8165–0.8864) and localization F1 (0.7491 vs. 0.2507 on VOC; 0.7694 vs. 0.1713 on SD; 0.3022 vs. 0.0862 on CelebA), while Table 4 documents the full trade-off up to λsyn=10000\lambda_{syn}=10000 (~27 dB).

3. Baseline Selection and Motivation Consistency

  • SepMark: SepMark performs whole-image deepfake detection (binary real/fake classification) and source tracing. It outputs no tamper localization masks and performs no image reconstruction, and is therefore not a baseline for manipulation localization or content reconstruction.
  • LIDMark: LIDMark is face-specific, relies on sparse 68-landmark displacement consistency rather than dense pixel-level segmentation masks, and does not reconstruct tampered RGB content. It is therefore applicable to watermark payload and BER comparison (Table 8), but incompatible with pixel-level mask F1/AUC and RGB reconstruction benchmarks.
  • Passive IML (Mesorch, PNF-Net, TruFor, ObjectFormer): Our primary competitors are proactive frameworks (EditGuard, OmniGuard, Imuge+, RecoverMark), not passive detectors. Nonetheless, we already evaluate three passive baselines including the ICLR 2026 SOTA RelayFormer to represent modern passive localization performance.
  • Motivation Consistency: We have refined the Introduction to align with Appendix A.1 and Table 5: while modern passive localizers improve distortion resilience (and classical methods degrade under JPEG), passive detectors rely on manipulation types seen during training rather than a universal, manipulation-agnostic prior, and still provide only an alarm without a rescue (lacking active ownership verification and content reconstruction). In contrast, existing general proactive methods lack principled algebraic invariants and suffer severe degradation under global distortions.

We thank you again for your thorough review and thoughtful questions. We hope these clarifications and manuscript updates resolve all remaining concerns and encourage a positive reconsideration of your final rating.

Best regards,

Authors

Thanks to the author's efforts, I raised my score from 2 to 3.

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

This paper studies robust tamper localization in proactively watermarked images. Both localized content edits and global post-processing, such as noise or JPEG compression, can disrupt the embedded protection, making it difficult to distinguish tampering from benign distortion. Within this existing proactive-forensics setting, the paper introduces continuous block-wise algebraic constraints during watermark embedding and derives syndrome residuals as a structured spatial prior. SyndromeMark combines this prior with RGB context for localization and employs a reconstruction network with a syndrome-consistency training loss for semantic inpainting. The experiments report improvements over several proactive baselines in some distortion settings, although the advantage is not uniform across all evaluated conditions.

Critical strengths

S1: Clear presentation and organization. The paper is generally well written. The progression from the linear CSF formulation to its practical limitations and the network-augmented framework is easy to follow, and the figures help explain the main information flow.

S2: Coherent system design. The integration of proactive embedding, ownership verification, manipulation localization, and semantic reconstruction provides a concrete end-to-end workflow. The combination of a structured residual with learned image context is a well-motivated design choice.

S3: Useful evaluation scope. The experiments cover multiple manipulation types and common post-processing distortions. Training on 5,000 synthetic images and testing on unseen natural-image datasets without fine-tuning is also an interesting setting for examining generalization.

Critical weaknesses

W1: The prior’s specific benefit is not sufficiently isolated. Figures 4 and 7 suggest that syndrome inputs help, but these qualitative ablations do not isolate the algebraic construction’s benefit from that of an additional protection-derived cue. The theoretical bounds are conditional, and their relevance to realistic tampering–distortion separation is not quantitatively demonstrated. Thus, attributing the robustness gains specifically to the algebraic design requires stronger evidence.

W2: Robustness gains need clearer qualification. The advantage is not uniform: on CelebA under JPEG, F1 is 0.4350 versus RecoverMark’s 0.7870 (Table 1). Meanwhile, stronger syndrome constraints improve the reported SNR but reduce protected-image PSNR from 32.36 to 26.98 dB (Table 2). Without clearer baseline settings, mask-threshold selection, and fidelity-controlled comparisons, the scope of the robustness advantage and its image-quality cost remain unclear.

Questions for authors

Q1: Under matched training and image fidelity, how does the syndrome prior compare quantitatively with RGB-only and alternative protection-derived cues, including distortion-only false positives? Can residual measurements show whether the theoretical separation conditions hold in practice?

Q2: Could you clarify baseline configurations and mask-threshold selection, and compare localization at matched protected-image fidelity? Please identify the distortion settings where SyndromeMark improves over strong baselines and those where it does not.

Reviewer_ZcL6OpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a proactive image forensics approach that combines continuous algebraic constraints with neural networks for manipulation localization and content reconstruction. CSF provides a syndrome-based formulation, while SyndromeMark uses it to guide localization and semantic reconstruction under post-processing distortions. Experiments assess the framework across several manipulation types and distortions, with ablation studies examining the role of the syndrome prior.

Critical strengths

The proposed framework builds on a solid algebraic foundation and combines it with neural networks for manipulation localization and semantic reconstruction. The connection between the continuous syndrome formulation and the network design is thoughtfully developed, resulting in a coherent and conceptually innovative approach to proactive image forensics.

Critical weaknesses
  1. Several baselines achieve higher localization F1 scores (Tables 1, 3–5), while SyndromeMark’s PSNR gains do not consistently extend to LPIPS (Tables 1 and 5). The explanation for moderate F1 also lacks threshold or boundary analysis. These results need more qualified claims about localization and reconstruction performance.

  2. The algebraic benefit for reconstruction is not directly validated. Section 5.2 provides neither an ablation of the syndrome-consistency loss nor measurements of output constraint errors. These are needed to establish whether the proposed constraint improves reconstruction beyond the network’s standard inpainting ability.

Questions for authors
  1. Could the authors clarify the evaluation protocol, particularly how localization thresholds are selected and how ground-truth masks for softly blended edits are constructed and binarized?

  2. Could the authors provide the specific PASCAL VOC and CelebA subsets, the number of test samples in each evaluation setting, and the editing tools used to generate the manipulations? Please also describe the synthetic image generation process and how the training and validation sets are partitioned.

1 comment
AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Revised Manuscript

Dear Reviewer ZcL6,

As the author-reviewer discussion period is approaching its end, we wanted to follow up briefly on our earlier response and revised PDF. We hope our clarifications and additional experiments have resolved your concerns, and we would warmly welcome you to join the discussion if you have any remaining questions or feedback before the deadline. If you find that your concerns have been satisfactorily addressed, we kindly invite you to consider updating your rating accordingly.

Best regards,

Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Reviewer ZcL6

Dear Reviewer ZcL6,

Thank you for your follow-up and for noting that most of your concerns have been addressed. Below we concisely clarify both points.

1. Ground-Truth Mask Generation

In our code, the adaptive high-percentile threshold within the face region uses the 99th percentile of the normalized mean absolute RGB difference, bounded below by 0.020.02, followed by 5×55\times 5 dilation, 3×33\times 3 erosion, and connected-component filtering. Because InstructPix2Pix is a latent diffusion editor that reconstructs the whole canvas through its VAE and U-Net, low-amplitude pixel drift occurs across the entire face. Using the 99th percentile (with the 0.020.02 floor) isolates the genuine localized semantic edit (such as scars or eyeglass frames) from background diffusion drift, which directly aligns with the 0.798%0.798\% average edited pixel area reported in our analysis.

2. Dataset Subsets and Sample Selection

  • Exact Splits: Both the PASCAL VOC images and the CelebA images were randomly sampled from the official training splits of PASCAL VOC 2012 and CelebA.
  • Why the Split Choice Causes Zero Data Leakage: Our entire framework (Phases 1–3) and the threshold calibration set were trained and calibrated exclusively on synthetic images generated by Stable Diffusion 2.1. No natural images or facial images from any split of PASCAL VOC or CelebA were ever used during training, validation, or calibration. Thus, the official training splits remain strictly unseen zero-shot test sets for SyndromeMark (and, if anything, are conservative against us since evaluated baselines were trained on natural and facial training distributions).
  • Sample Selection Criteria: For PASCAL VOC, test images were randomly sampled from the training split. For CelebA, after running InstructPix2Pix on randomly sampled faces, we retained cases where the diffusion editor produced clean, visually realistic localized edits (excluding failed or globally distorted generations, consistent with our realistic-forgery assumption in Section 3.2). Notably, these retained edits are very small (0.798%0.798\% pixel area on average) and therefore remain challenging for localization.

We have updated Appendix C.2 in the revised manuscript to record the exact 99th percentile rule and the official training-split sampling details. We hope these final clarifications fully resolve your remaining questions, and we kindly invite you to reconsider your final rating.

Best regards,

Authors

Dear reviewers,

As the discussion period approached the end, please take a moment to read the authors’ responses and share your thoughts on any remaining points of discussion. If the discussion has changed your assessment, please update your review accordingly.

Thank you for your time and contributions!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.