acceptodds
Under review as a conference paper at ICLR 2027

Doctor Trigger: A Framework for Release-Bound Face Verification

Abstract

Face swapping and generative editing enable realistic facial manipulation, while online redistribution complicates publisher oversight of circulating face images. Existing approaches establish provenance, detect manipulation, recover embedded watermarks, verify image integrity, or disrupt unauthorized face manipulation. These capabilities do not necessarily establish consistency with a specific publisher-enrolled release, limiting the evidence available for verification and human review. To address this gap, we introduce Doctor Trigger, a framework for verifying circulating faces against private publisher-held release records. Before publication, the publisher embeds a weak keyed Fourier-phase signal in each face while preserving visual fidelity and stores its spectral reference. Afterward, verification compares the circulating face with that reference and uses key projection to measure retention of the release-specific signal. Therefore, relation and key evidence jointly help publishers identify suspicious copies for review, potentially reducing the circulation of manipulated images. Experimental results across FF++, DFDC, and Celeb-DF show that spectral agreement separates tested face swaps and generative manipulations from enrolled releases, while the keyed signal distinguishes published versions from their unreleased originals under the evaluated key-preserving benign transformations. Doctor Trigger complements provenance metadata with release-specific face verification and record-linked evidence for further assessment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Release-bound face verification, Media provenance, Proactive image forensics, Fourier phase modulation, Reference-based verification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zNtMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript introduces doctor trigger, a framework designed for publishers to verify whether a circulating face image corresponds to a specific, official release. The system works by embedding a weak, keyed phase displacement in the Fourier domain of the image before publication. The authors evaluate their framework on multiple deepfake datasets under various benign processing channels.

Critical strengths
  1. The problem of deepfake provenance and publisher-side verification is a pressing real-world challenge. The paper targets a very practical scenario where publishers need to distinguish between an official release, a leaked pristine original, and a maliciously manipulated image.
  2. The authors provide extensive experiments across multiple datasets and evaluate the system against various common transmission distortions (e.g., JPEG compression, blurring, social media reposting).
Critical weaknesses
  1. The core verification mechanism relies almost entirely on handcrafted Fourier-domain operations, including phase modulation, spectral agreement, and keyed projection. The only task-specific learned component is a lightweight two-hidden-layer MLP trained on handcrafted spectral statistics, while the image-only detector serves as an auxiliary review module. Consequently, the manuscript emphasizes system integration and verification workflow design rather than introducing a substantive new machine learning methodology. Its technical contribution appears relatively limited for ICLR.
  2. Embedding imperceptible information through Fourier-phase modulation is a well-established technique in digital watermarking, dating back several decades. Although the proposed release-bound verification setting is practically relevant, the underlying signal embedding and verification mechanisms remain closely related to classical reference-based watermark authentication. The paper does not sufficiently establish what fundamental algorithmic advances are introduced beyond adapting these existing techniques to face-image provenance verification.
  3. The framework assumes that the query image can be associated with the correct enrollment geometry, which significantly limits its applicability to unconstrained internet circulation. Although the authors explore a limited cropping scenario, the results themselves reveal substantial fragility: under 90% center cropping, the face-swap rejection rate drops to only 6% despite known crop geometry and additional alignment correction (Table 4). In practice, circulating images frequently undergo unknown cropping, padding, rescaling, and geometric distortions. The lack of a reliable mechanism for handling such transformations remains a major limitation for a practical release-verification framework.
  4. The results in Appendix K reveal substantial vulnerabilities to diffusion regeneration and targeted adversarial perturbations. More importantly, the phase-ring transplantation attack in Appendix K.2 demonstrates that release-specific spectral evidence can be transferred to manipulated facial content, with 84% of attacked FF++ queries receiving a confirmed verdict under the learned joint verification policy. Although this attack assumes access to the enrolled active-bin mask, the results raise concerns about whether the proposed spectral evidence is sufficiently tied to the actual facial content. Such vulnerabilities undermine the reliability of the verification mechanism, particularly because incorrectly confirmed manipulations may bypass the proposed human-review workflow.
  5. Although the manuscript emphasizes a practical publisher-side human-in-the-loop workflow, several essential components, including multi-face association, public credential integration, and human-review effectiveness, remain conceptual or unevaluated. Consequently, the experiments primarily demonstrate the effectiveness of spectral comparison under controlled conditions rather than validating the complete deployment-oriented framework claimed by the paper.
Questions for authors

See weaknesses.

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

This paper introduces a watermarking method that allows a publisher to inject noise into a facial image before release. The noise is driven by a unique signature that is recorded. The publisher offers a verification service for those who want to check whether a given facial image originated from an image into which the publisher injected noise. The noise signal survives image-editing operations such as face swapping. In addition, the signal should survive benign image processing operations that may occur during normal image delivery. Experiments have been conducted to empirically demonstrate the effectiveness of verification.

Critical strengths
  • The intention to combat face-related misuse is of societal importance.
Critical weaknesses
  • The presentation of this paper is poor. Many passages (e.g., Lines 141–144) are incomprehensible. While the problem setting can be easily described, the wording used in this paper makes it difficult to understand. The reviewer recommends adopting standard terminology used in image watermarking unless the authors believe there is a substantial difference.
  • There is a serious issue regarding the operationalization of the proposed method. In particular, when a person is interested in verifying a facial image, which verification service provider should they approach?
  • While image processing operations are considered to improve robustness in the real world, face detection and cropping can change the signals drastically. However, the proposed method seems to assume that the input image is always cropped perfectly and consistently during noise injection and verification.
  • The authors should carefully review the paper and include proper citations (e.g., C2PA, StarGAN, etc.).
Questions for authors
  1. What is the threat model?
  2. Could you give a real-world example of how the proposed pipeline would work?
  3. Given a facial image to be verified, which verification service provider should be approached?
  4. How does the proposed method perform under adaptive attacks?
2 comments

Thank you for the response. While it addresses some concerns, the weaknesses still outweigh the strengths of this paper. Considering other reviewers' comments, I would increase the score to weak reject.

Dear Reviewer LAKX,

Thank you for reading our response; we are glad that you have raised your score.

However, it is not clear to us which specific weaknesses you feel still outweigh the strengths. We have responded point by point, on the facts and with reference to the paper and new experiments, to each of the four weaknesses and four questions you raised: terminology and presentation (W1), which verification service to query (W2, Q3), cropping (W3, Table R6), citations (W4), the threat model (Q1), a real-world example (Q2), and adaptive attacks (Q4, Table R4).

If you are referring to the limitations in the paper, such as the attack results in Appendix K and the cropping experiment in Table 4, we would like to note that we reported these ourselves, following the ICLR Code of Ethics, which asks researchers to “provide full disclosure of all pertinent system capabilities, limitations, and potential problems,” as mentioned in our earlier response. We could have chosen not to report these results, but we believe that honestly disclosing a method’s limitations and applicability boundary is the responsible thing to do; if proactive disclosure is treated as a weakness, authors will stop reporting such results, which does not serve the research community, especially on safety-related topics.

Moreover, during the rebuttal we added experiments showing how to address these potential weaknesses: an out-of-ring low-frequency check, which requires no training, reduces phase-ring transplantation to 0% (Table R2); under the full policy, the other attacks in Appendix K do not meaningfully raise false confirmation (Table R4); and when a circulating photo is cropped or reframed, the system outputs unavailable rather than a false confirmation (Table R6).

We are concerned that if the remaining weaknesses cannot be traced to specific, verifiable points in the paper, the assessment may come to rest on subjective impressions rather than on the paper itself. If you consider any of our responses incorrect or insufficient, we would therefore be grateful if you could point out which one and why; we would be glad to address each point before the discussion period ends and to improve the revision accordingly.

Reviewer_7iMxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a release-bound face verification framework that combines a private spectral reference with a weak keyed Fourier-phase signal. The spectral reference measures consistency with an enrolled face, while the keyed signal helps distinguish the published version from its original. The authors introduce both direct threshold-based verification and a learned record-based model, along with a human-review workflow. Experiments on FF++, DFDC, and Celeb-DF evaluate face manipulation detection, release-key discrimination, robustness to image processing, and computational cost. The results show strong performance under several tested conditions, although robustness degrades under stronger processing and adaptive attacks.

Critical strengths
  1. The separation of face consistency and release-key evidence is well motivated and provides a useful framework for publisher-side verification.
  2. The authors evaluate multiple datasets, manipulation methods, processing channels, and reference-based baselines. The inclusion of operating-point analysis and adaptive attacks is valuable.
Critical weaknesses
  1. The main components, including Fourier-phase watermarking, private-reference comparison, and keyed verification, are closely related to existing techniques. While their combination is useful, the paper does not convincingly establish a substantial technical advance over prior reference-based or proactive watermarking approaches.
  2. The phase-ring transplantation attack is particularly concerning: the learned joint policy confirms 84% of attacked FF++ queries. This directly undermines the proposed release-bound verification objective. Moreover, the attack evaluation assumes access to the enrolled active-bin mask, leaving uncertainty about practical exploitability.
Questions for authors
  1. What is the main technical advantage of the keyed phase signal over existing content-dependent watermarking methods when both have access to private enrollment records? Could the authors provide a more directly comparable evaluation?
  2. How does phase-ring transplantation perform when the attacker has access only to the public released image, without the private active-bin mask? This seems important for understanding the practical severity of the reported vulnerability.
  3. Could the authors report final joint verification results under the adaptive attacks, rather than primarily component-level AUCs?
3 comments

Dear Reviewer 7iMx,

We apologize for following up once more, and we fully understand how busy this period is. As the discussion period is about to close, we would like to highlight the single result most relevant to your main concern: phase-ring transplantation succeeds equally without the private mask (Table R1), and an out-of-ring low-frequency check, which requires no training, reduces it to 0% on all three datasets (Table R2).

Even a one-line reply on whether this addresses your concern would be very helpful to us and to the AC. Thank you for your time.

Dear Reviewer 7iMx,

We sincerely apologize for following up yet again, and we truly understand how busy this period is. We have been very encouraged that you found the separation of face-consistency and release-key evidence well motivated, and we took your questions to heart: the mask-free transplantation and the low-frequency check (Tables R1 and R2) were run specifically in response to your Q2, which helped us understand the practical severity of this vulnerability.

Your perspective genuinely matters to us. As the discussion period is closing very soon, we would be deeply grateful if you could take a look at our response and let us know whether it addresses your concerns, or what still remains unclear. If you have any further questions, we will respond as promptly as possible.

Thank you very much for your time and for your careful review, and apologies again for the repeated messages.

Sincerely, The Authors of Submission 6431

Dear Reviewer 7iMx,

We are sorry to reach out once more. We know your time is limited, and we truly appreciate the effort you have already put into reviewing our paper. Before the discussion period ends, we would be very grateful to hear your thoughts on our response, in particular on the mask-free transplantation and the low-frequency check we added for your Q2 (Tables R1 and R2). Any question you have, we will answer right away.

With sincere thanks, The Authors of Submission 6431

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.