acceptodds
Under review as a conference paper at ICLR 2027

Image Watermarking with Memory across Reissuance

18 views

Abstract

Publishers often revise and reissue images, yet valid watermarks may persist on superseded releases or unapproved edits. Even a current release can contain regions last authorized at different revisions. We introduce and formalize reissuable provenance: determining an image's current, superseded, or unissued status and recovering each region's last authorized revision (its generation), with abstention when evidence is insufficient. Generational Stratigraphy Fields (GSF) combine watermark evidence with an authenticated live reference to the current release, enabling replacement or retirement without rewriting circulated copies. Each reissue freshly embeds new generation labels in approved regions and inherited labels elsewhere, preserving history without requiring old watermarks to survive successive edits. Our complete-output reissuance evaluation follows 480 histories through four releases, retaining each AI editor's entire output. GSF achieves 99.67% accuracy in classifying issued images as current or superseded and 94.09% recall of unissued edits. After three reissues, one generation map provides time and location: it recovers regional labels with 93.13% median accuracy and localizes the latest approved revision with 92.86% IoU. Regional labels remain readable under tested delivery transformations that impair issued-state recognition. Watermarks can thus express who issued an image, which release is authoritative now, and when each region was last authorized.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image watermarking, Content provenance, Image reissuance, Live-head verification, Regional edit history, AI image editing, Content authenticity

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TCKyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies reissuable image provenance: distinguishing current, superseded, and unissued copies, with abstention when evidence is insufficient, and recovering regional last-authorized-revision labels. GSF combines a trusted current-issue reference with image-embedded evidence, refreshing the entire field at each issuance while preserving inherited labels. Two-bit keyed fingerprints and a candidate-equivariant spatial reader recover regional history. Across 480 lineages containing four issued versions, the method reports 99.67% clean issued-state accuracy, 94.09% unissued-edit recall, and 93.13% median regional accuracy after three reissues.

Critical strengths
  1. The paper clearly distinguishes current authorization from historical authenticity and presents a coherent problem formulation and reissuance protocol.
  2. The evaluation covers complete editor outputs, multiple editors, and lifecycle events. Reader ablations and explicit failure reporting make the empirical analysis informative.
Critical weaknesses

W1. Evidence against stronger task-matched alternatives remains limited. Table 3 scores global-message carriers as constant regional maps, while the WAM adapter refreshes only committed field cells and relies on older marks for inherited history. This limits the evidence for GSF’s advantage over an optimized full-field-refresh regional baseline. Appendix M.7 also shows that correctly identifying an issued version allows exact map reconstruction from full registry records. Avoiding full registry reads is valuable, but stronger comparisons under matched registry-access assumptions are needed to establish GSF’s incremental advantage.

W2. Scalability to longer revision histories remains unvalidated. The end-to-end histories contain only four issued versions, where direct two-bit labels can represent all versions without codebook collisions. However, growing candidate sets are a central motivation for the keyed design. Appendices K and L analyze collision and width trade-offs but do not demonstrate longer-history recovery. Candidate-equivariance establishes interface compatibility, not recovery accuracy as unique anchors become scarcer, especially for small or fragmented older regions.

W3. Issued-state recognition deteriorates sharply under some benign delivery transformations. Table 53 reports CURRENT retention of 35.0% under JPEG quality 50 and 2.5% under composite delivery proxy B, versus 98.5% and 99.4% for TrustMark-Q with the same protocol. Readable regional maps do not eliminate this issued-state error. Appendix E explains these failures, and the paper includes limited operating-point comparisons, but does not quantify the trade-off between benign-delivery tolerance and false acceptance of unissued edits across operating points under these delivery conditions.

Questions for authors
  1. Regarding W1: Can the authors evaluate an optimized full-field-refresh regional baseline at comparable payload and image-quality budgets, explicitly distinguishing deployments with and without full registry-history access?
  2. Regarding W2: How do regional accuracy and mIoU change beyond four issued versions, especially for small or fragmented older regions? How does the two-bit design compare with wider codes under a fixed total carrier budget?
  3. Regarding W3: Can the authors quantify the issued-state retention/false-acceptance trade-off under these delivery transformations and clarify the intended deployment envelope?
5 comments

Dear Reviewer TCKy,

Thank you for your careful and constructive review, and for finding the formulation clear, the protocol coherent and the failure reporting informative. We ran each experiment your questions ask for on all 480 lineages.

Our response in brief (Tables R5–R7 are new; R1–R4 are in our reply to Reviewer qA8C.)

  1. Q1, regional baseline. Without full registry-history access, a full-field-refresh WAM baseline, tuned in its favor, stays 30.3 / 43.7 / 44.8 mIoU points below GSF on inherited history (depths 1–3) at matched image quality, and 13.5 / 31.6 / 35.8 points below when it spends about 9 dB more distortion (Table R5). With full access, the map of an issued image is a lookup for any carrier (App. M.7); our reply to Reviewer qA8C answers that half in detail.
  2. Q2, longer histories. From four to eight versions, Map acc / mIoU go from 93.1 / 79.0% to 86.9 / 59.0% (Table R6). Cells without a unique anchor are still read at 84.8%: recovery accuracy carries over. The loss sits in small fragments, and beyond the declared depth a three-bit code can do better, as Sec. 4.2 states.
  3. Q3, harsh delivery. Sweeping one threshold (C_cc) from its default to 12 raises retention under JPEG 50 / Platform B from 35.0 / 2.5% to 93.8 / 90.4%, and the unissued edits accepted from 0.5 / 0.0% to 3.4 / 2.1% there and from 2.5% to 8.4% on clean copies. TrustMark-Q retains 98.3 / 99.2% and accepts 24.0 / 26.1% (Table R7). The intended envelope is the 18 conditions of App. E.3; JPEG 50 and Platform B lie outside it.

Q1. Full-field-refresh regional baseline, with and without full registry-history access

The paper's WAM row rewrites only committed cells, as you note (coexisting WAM messages interfere, App. M), whereas GSF "rewrites every field cell" at each reissue (Sec. 4.2). No regional watermark we know of defines a reissue step, so we built the baseline from the closest released one, WAM, with every choice in its favor (≈ 47 GPU-hours): all 256 cells re-embedded at every reissue; GSF's payload per cell; two quality budgets, matched to GSF and at WAM's released strength; the reader most favorable to it, selected on the test lineages, and status thresholds tuned on a 2,268-point grid.

Table R5. Inherited regional history read from the copy (mIoU, %)

Regional watermark (first-issue PSNR)Depth 1Depth 2Depth 3
WAM, committed cells only (Table 3, Sec. 5.5; 37.8 dB)49.132.424.3
WAM, full-field refresh, matched quality (46.7 dB)63.643.634.2
WAM, full-field refresh, released strength (37.8 dB)80.455.743.2
GSF (46.7 dB)93.987.379.0

Without full registry-history access, full refresh helps the baseline, as you expected, and makes it a working one: tuned, it reaches CURRENT / SUPERSEDED 90.9 / 95.7% at released strength (GSF: 99.5 / 99.8%; Table 3, Sec. 5.5). Yet on inherited history it stays below GSF at both budgets, for a structural reason: on older regions of at most 16 cells its median IoU is 0 (GSF: 40.0% and 66.7% for regions of 1–4 and 5–16 cells), also on control images with no earlier mark to interfere. A region-level message needs a large contiguous area to decode; GSF's keyed field is addressable cell by cell.

With full registry-history access, any carrier replays the map of an issued image exactly from the records, as the paper states: such a deployment "needs no field on issued images" (App. M.7). The comparison reduces to identifying the issue: 94.1% / 97.1% for the baseline at the two budgets, its tolerance widened in its favor, against 99.97% for GSF. For an unissued copy the target is its parent's field (Sec. 3), so a replayed record is right whenever the parent is identified. Our reply to Reviewer qA8C measures what that takes: a hash-keyed lookup with the complete record matches 0.1–42.0% (by tolerance) of 960 locally edited copies, each match certifying an unissued copy as an issue; GSF recognizes 99.8% of them as unissued and reads their untouched regions at Map acc 93.7% / mIoU 79.4%, against 93.6 / 79.7% on the issued image.

Q2. Longer histories; small or fragmented older regions; two bits against wider codes

The submission stops at four versions (App. F.2). We extended every lineage to eight (1,920 further reissues; ≈ 23 GPU-hours) with writer, decoder and reader frozen.

Table R6. Regional history from four to eight issued versions (medians, %)

Issued versions4 (paper)5678
Map acc93.191.390.088.186.9
mIoU79.072.567.162.859.0
Ambiguous cells (no unique anchor): measured / App. L58.1 / 57.868.8 / 68.476.3 / 76.382.5 / 82.286.9 / 86.7
Map acc on ambiguous cells88.387.486.985.784.8
Three-bit code, error-free ceiling: Map acc / mIoU90.6 / 76.090.3 / 73.890.3 / 72.189.9 / 70.489.9 / 69.4

(i) Beyond four versions. Ambiguous cells grow as App. L predicts (Table R6), yet are still read at 84.8% (88.3% at four versions). Labels at least four issues older than the current one, which a direct two-bit code cannot distinguish from a later label (App. K), keep a median IoU of 75.0 / 61.5 / 50.0 / 50.0% at five to eight versions.

(ii) Small or fragmented older regions. The loss follows fragment size. At five to eight versions, older-label cells are read correctly at 42.9% when isolated and 62.2% in fragments of 2–4 cells (together 15.5% of these cells), against 85.3% and 92.4% in fragments of 5–16 and ≥ 17 cells; at 5–16 cells a fragmented region scores below a contiguous one (median IoU 60.0 against 71.4%). The paper names this mechanism: "because older slices become small and fragmented" (Sec. 5.3).

(iii) Two bits against wider codes. Under the same 320-slot budget a three-bit code buys unique anchors with 106 coarser cells instead of 160 (App. K). Within the declared depth, the measured two-bit field exceeds even the error-free ceiling of that code (Table R6, last row), and the four- and eight-bit ceilings stay below the measured field at every depth. Beyond it the three-bit ceiling takes over, in mIoU from five versions and in accuracy from six. Sec. 4.2 states both: "two bits maximize the worst-case expected number of uniquely identifying cells over d ∈ {1,2,3} … deeper histories gradually favor wider codes".

Q3. Retention against false acceptance under JPEG 50 and Platform B; the intended envelope

Your numbers are right, and so is your remark that a readable map does not remove this error: under JPEG 50 the map is unchanged (App. E.3); the status is lost because the default rule reads any disrupted cluster as evidence of an unissued change. But retention is one coordinate of an operating point; the other is what the system accepts under the same conditions.

The trade-off. With every other rule fixed (Table 10, App. B.5), we swept C_cc, the size a disrupted cluster must exceed to count as such evidence, and re-decided all 61,440 delivery queries at 10 values. TrustMark-Q's delivered unissued outputs were decoded under the same protocol (≈ 10 GPU-hours): from the images saved in the paper's run for 398 lineages and, for the other 82, from lineages regenerated with the same instructions, which moves its retention by one lineage per condition from the paper's 98.5 / 99.4% (Table 53, App. M.6).

Table R7. CURRENT issues kept ↑ / unissued edits accepted as issues ↓ (%)

Operating pointJPEG 50Platform BClean
GSF, C_cc = 1 (the paper's default)35.0 / 0.52.5 / 0.099.0 / 2.5
GSF, C_cc = 684.8 / 2.268.3 / 0.899.4 / 6.1
GSF, C_cc = 1293.8 / 3.490.4 / 2.199.4 / 8.4
TrustMark-Q, same protocol98.3 / 24.099.2 / 26.199.8 / 29.6

TrustMark-Q keeps 98.3 / 99.2% because its global message survives, and so accepts 24.0 / 26.1% of the unissued edits there. C_cc = 12 is the first swept value at which retention under JPEG 50 passes 93%; the price is paid on clean copies (false acceptance 2.5% → 8.4%, OPEN recall 93.1% → 82.7%). We show it as a point on the curve, not as a recommended setting.

The intended envelope is that of the default point, whose thresholds were placed on development data before the evaluation: the 18 conditions of App. E.3 in which all three statuses stay above 93% (clean; JPEG down to quality 75, including double JPEG 75; downscaling to 256 px; blur up to σ = 1.0; darkening to ×0.85; contrast ×0.80). JPEG 50 and Platform B lie outside it, and, as the sweep shows in hindsight, no swept value keeps all three statuses above 93% on clean copies and under either of them at once. Outside it, a damaged issue and an unissued edit can produce the same evidence (App. E.5), and the threshold moves errors from one to the other without separating them.

Outside the envelope, two facts. None of 30,720 issued-image queries over 32 conditions is given the other issued status, at any swept value; an issue that fails reads OPEN or UNKNOWN (Sec. 5.4). Unissued outputs are accepted at 2.5% on clean copies and at most 2.6% under any transformation at the default point.

In closing. The experiments leave the paper's conclusions in place: "locating the latest change is not keeping the history behind it" (Sec. 5.5) holds against the stronger baseline; two bits remain the right width for the declared depth, and three bits can do better beyond it (Sec. 4.2); under harsh delivery the cost falls on retention (Sec. 5.4). In the revision, Table 3 (Sec. 5.5) will report the full-field-refresh baseline at both budgets, with its construction and tuning detailed in App. M; the eight-version results go to App. G and L, the sweep to App. D and E, where the App. E.3 sentence on unissued outputs becomes these two facts. We would be grateful if you would reconsider your assessment.

Sincerely, The Authors

AuthorsOpenReview ↗

Follow-up on the additional experimental results

Dear Reviewer TCKy,

Thank you again for your helpful questions. We would appreciate any further feedback on the full-field-refresh baseline, longer-history evaluation, and delivery trade-offs reported in Tables R5–R7.

In particular, could you please let us know whether these results address your main concerns, or whether any points remain unresolved? We would be happy to provide further clarification during the remaining discussion period.

Thank you for your time.

Best regards, The Authors

Reviewer_W5b1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for tracking the provenance and regional editing history of images across multiple rounds of reissuance. The method uses Generational Stratigraphy Fields (GSF) to encode regional version information into images and combines this information with an external trusted live-state registry to determine the current status of an image and its regions. The framework divides images into a spatial grid and uses keyed fingerprints and spatial decoding to recover regional history. The evaluation includes 480 COCO root images, four generations of reissuance, four image editors, and 2,400 editing calls. The paper reports 99.67% accuracy for whole-image state classification, 93.13% regional generation accuracy, and 78.96% mIoU for regional history recovery. The results suggest that the proposed framework is promising for tracking image editing histories across multiple generations.

Critical strengths
  1. A practically relevant problem and a useful step beyond existing approaches. The paper addresses a practical problem: tracking both the current authorization status of an image and the last authorized revision of each region across multiple rounds of reissuance. These two aspects are often treated separately, whereas the proposed framework combines them through an external trusted live-state registry and an embedded regional history field. This provides a useful framework for reasoning about image provenance and revision history in multi-generation reissuance workflows.

  2. A more realistic evaluation of AI-edited images. The evaluation retains the complete outputs returned by AI image editors rather than pasting the edited regions back into the original images. This is an important design choice because AI editors may also reconstruct nominally unmodified regions, making the resulting images different from the original beyond the explicitly intended edit areas. By evaluating complete editor outputs across multiple generations, the paper considers a more realistic and challenging setting for image reissuance than an evaluation that assumes the background remains pixel-wise unchanged.

Critical weaknesses

I do not identify a critical weakness that independently warrants rejection. The questions mentioned in the next section could benefit from further clarification or additional analysis, but I consider them non-blocking concerns rather than fundamental flaws.

Questions for authors
  1. Robustness against active watermark attacks

The paper provides extensive evaluations under various image post-processing conditions. However, it remains unclear how the method performs against active attacks specifically designed to remove, disrupt, or forge watermarks. Could the authors clarify the scope of the current robustness evaluation and discuss the method’s limitations against such attacks? Additional experiments, if feasible, would further strengthen the evaluation.

  1. Overlap among approved editing regions across generations

The paper evaluates regional history recovery after multiple rounds of image reissuance. However, it is unclear how much the approved editing regions overlap across different generations and whether this spatial overlap affects the reported regional history recovery performance.

Since each region records the most recent generation in which its modification was authorized, the overlap pattern across generations may affect the final distribution and spatial structure of regional generation labels, potentially influencing the difficulty of the recovery task.

Could the authors report overlap statistics between approved editing regions across adjacent and non-adjacent generations, and clarify whether the reported results are sensitive to these overlap patterns? If feasible, an analysis of performance under different overlap levels would help establish the stability of the method across different regional editing patterns.

My questions have been adequately addressed by the authors' responses and additional experiments. I therefore maintain my original rating of 3.

4 comments

Dear Reviewer W5b1,

Thank you for your positive and careful review. We are glad that you found the problem practically relevant and valued the external trusted registry for live authorization, the regional history carried in the image, and the evaluation on complete editor outputs. Your two questions point at what the submission left implicit; we have now measured it.

Our response in brief (Tables R8–R9 are new; R1–R7 are in our other two replies.)

  1. Q1, active attacks. The submission tests no attack designed against the mark. Under removal, disruption and forgery by an attacker without the key (Table R8), an issue can be made unverifiable. An unissued copy, however, is accepted as an issue in 0 of 2,400 cases after removal, 0.0–3.3% after a transplant of the issue evidence and 0.4–8.3% after a copy of the whole mark (0.0–2.5% with no attack; global-message watermarks under the same copy attack: 96.5–99.8%). Q1 (iii) lists four limitations.
  2. Q2, overlap. Approved regions overlap, about as much for non-adjacent as for adjacent issues (median IoU 0.43 against 0.45 / 0.44). The results are sensitive to overlap in one identified way: it shrinks what remains of the earlier regions, and small remnants are read less accurately. Across quartiles of observed overlap, Map acc varies by at most 0.6 points and mIoU falls by 4.0; under controlled overlap they fall by up to 2.3 and 15.6 points (Table R9).

Q1. Robustness against active attacks: scope, new experiments, limitations

(i) Scope. The submission evaluates benign delivery (31 transformations; Sec. 5.4, App. E), complete-output edits by four editors (Sec. 5.1) and queries from outside the lineage (App. N). It contains no attack designed against the mark, and you are right to ask: the paper only lists "forged marks" in the "wider input" that a deployed verifier meets (App. N). Its claim concerns authorization, not unremovability: historical bitmaps "cannot become a new issue unless the issuer authenticates the issuance transaction" (App. A.4). How often the verifier's answer agrees with this under attack was not measured; we measure it now.

(ii) New experiments. The attacker has the public writer and decoder, the current issue I3 and unissued copies, but not the issuer's key; the verifier and its frozen operating point (Table 10, App. B.5) are unchanged (all 480 lineages, about 13,000 attacked copies). We report two failures separately: an issue loses its status ("the prediction authorizes nothing", Sec. 3), or a copy the issuer did not issue is accepted as an issue (read CURRENT or SUPERSEDED). For removal, VAE re-encoding, diffusion purification at three strengths and a bilateral filter are applied to I3 and to an editor's output of I3; a 3×3 median filter is a mild control. For disruption, a third-party watermark (TrustMark-Q, VINE-R or WAM) or a second GSF mark under another key is embedded over I3. For forgery, the attacker transplants from I3 the blocks of the 96 cells that carry lineage and issue evidence, or copies its whole mark (all 256 cells decoded and re-embedded), onto four kinds of unissued copy (an editor's complete output of I3 or of an earlier issue; I3 with 40% of its cells replaced; I3 with one object regenerated).

Table R8. The two failures under attack

AttackIssues still CURRENTUnissued copies accepted
None (reference)475 of 48034 of 1,920 (0.0–2.5% by kind)
Removal (five operators)1 of 2,4000 of 2,400 editor outputs
3×3 median filter96.3%–
Third-party watermark over the issue90.6–98.8%–
Second GSF mark under another key0 of 480–
Forgery: transplant of the evidence cells–32 of 1,920 (0.0–3.3%)
Forgery: copy of the whole mark–112 of 1,920 (0.4–8.3%)
Copy attack on VINE-R / TrustMark-Q / OmniGuard–97.9 / 99.8 / 96.5%

An issue that loses its status reads OPEN or UNKNOWN; none was read SUPERSEDED. For an unissued copy, the verifier's default is not to authorize, so destroying evidence takes a status away and, in the 2,400 cases tested, created none (12 of 480 editor outputs are accepted with no attack). A transplant does not raise acceptance overall, because the copy's field stays as disrupted as the edit left it (OPEN in 86.7–97.5%). Copying the whole mark raises it to 0.4–8.3% (51.3–73.3% UNKNOWN, 18.3–48.3% OPEN): the issue evidence is keyed and "bound to a descriptor of the issued content" (Sec. 4.2), so the image passes only if its content's descriptor stays within the check's tolerance. Every accepted forged copy reads CURRENT; without an attack, the accepted outputs of an earlier issue read SUPERSEDED.

For the last row's global-message watermarks (App. M), the copy attack reads the message from each watermark's own current issue and re-embeds it into its own unissued output (480 editor outputs each; GSF on such copies: 7.5%). Nothing checks their message against the content.

Given a content check like GSF's, the three global-message watermarks' acceptance falls to 2.9–5.4%, at the price App. M.4 reports: "CURRENT accuracy under an issue's own head falls to 67.4–79.9%" (GSF: 475 of 480).

(iii) Limitations under active attacks.

  1. The mark can be removed by regeneration through a VAE or a diffusion model, as invisible watermarks in general are (Zhao et al., 2024; Saberi et al., 2024; App. N), by a bilateral filter and by a second GSF mark. The attacked issue cannot then be told from an unissued edit (App. E.5); the registry and the other copies are untouched. GSF is designed to limit the acceptance of copies that were not issued; it does not keep a copy of an issue verifiable against an adversary.
  2. The content binding is a mitigation, not a cryptographic binding: the tolerance of the check lets benign delivery pass, and 0.4–8.3% of unissued copies with a copied mark pass too.
  3. Local evidence must extend beyond one cell (Table 10, App. B.5): unissued content pasted over 5–40% of the field, with the evidence cells intact, is accepted in 3.3–5.8% (our reply to Reviewer qA8C, Table R4), and a paste confined to one field cell (0.4% of the image) in 34.4% (63.3% UNKNOWN).
  4. Not evaluated: key compromise; attacks optimized through the public decoder; repeated queries to the verifier.

Q2. Overlap among approved regions across generations

Statistics. A fixed rule derives each region from the editor's return (App. C.3), and the regions are large (medians 147, 147 and 152 of 256 cells at depths 1–3; App. G.1), so they overlap by necessity. Over the 160 field cells of the 480 lineages, the median IoU (IQR) is 0.45 (0.30–0.61) and 0.44 (0.30–0.59) for the adjacent pairs, 0.43 (0.30–0.55) for the non-adjacent pair, and 0.38 (0.27–0.47) between regions of different lineages.

Sensitivity (observed overlap). From the lowest to the highest quartile of overlap (mean pairwise IoU 0.13–0.80), Map acc is 93.2 / 92.9 / 92.7 / 92.8% and mIoU 80.8 / 79.3 / 78.2 / 76.8%. The fall is on labels 1 and 2 (IoU 72.7 → 60.8% and 81.8 → 73.3%); labels 0 and 3 do not fall. With more overlap, more of regions 1 and 2 is relabeled by a later issue: the cells still carrying label 1 or 2 fall from 61.5 to 32.6 per lineage. The cells of labels 0–2 then lie more often in small fragments, which are read less accurately (60.0 / 70.9 / 88.3 / 95.2% in fragments of 1, 2–4, 5–16 and ≥ 17 cells). Given the same distribution of fragment sizes, the decline in the accuracy of these cells across quartiles (91.3% → 89.1%) becomes a 0.8-point spread without a trend.

Controlled overlap. To vary only the overlap, we kept fixed the editor output used to construct each lineage's I3 and changed only its three approved regions: each became a band of 4 of the 16 grid lines that covers 0, 25, 50, 75 or 100% of the previous one. The frozen writer wrote the field; the frozen verifier and reader read it. With the original regions, this pipeline reproduces the paper's I3 bit for bit on all 480 lineages.

Table R9. Controlled overlap on the same images (%)

Share of the previous region covered0%25%50%75%100%
Lines of regions 1 and 2 keeping their label (of 4)43210
Map acc (mean)98.297.596.995.997.8
mIoU (mean)96.594.491.280.995.8
Newest label read correctly98.498.498.598.699.1
Labels 1 and 2 read correctly98.496.793.480.8–

Overlap does move the metrics. As what remains of regions 1 and 2 narrows from four lines to one, Map acc falls from 98.2% to 95.9% and mIoU from 96.5% to 80.9% (a one-line strip is all boundary); at 100%, where nothing remains of them, both return to 97.8% and 95.8%. The loss sits on those strips (last row); the newest label is read at 98.4–99.1% throughout. Writing the same fields on the unmarked source images changes Map acc by at most 0.1 points. The bands are simpler than the editors' regions, hence the higher values.

Why. Overlap decides which label a cell ends with and how much of an older region remains, not how a cell is written: "The logical update is local, but the physical write is complete" (App. B.2). In the paper's lineages, cells of the newest label approved once, twice and three times are read correctly at 94.7 / 95.0 / 95.3%. Small remnants are read less accurately and weigh more in IoU (App. G.2); depth acts in the same way (Sec. 5.3; our reply to Reviewer TCKy, Table R6).

In closing. An attacker without the key can make an issue unverifiable; in Table R8, GSF accepted at most 8.3% of unissued copies passed as issues. Overlap acts through the size of what remains of earlier regions. The revision will state scope and limitations in Sec. 5.4, add the threat model and Table R8 as an appendix and the overlap analysis to App. G.1. We would be grateful if you would consider raising your assessment.

Sincerely, The Authors

Reviewer_W5b1OpenReview ↗

My concerns have been addressed and I will maintain my original score

Thank you for the responses and the additional experiments.

Regarding my first concern about robustness against active watermark attacks, the authors have added experiments on watermark removal, disruption, and forgery. The authors distinguish between two types of failures: issued images becoming unverifiable and unissued images being falsely accepted. While removal attacks mainly affect the verifiability of issued images, the forgery experiments show that some unissued images can still be falsely accepted, with the rate depending on the attack setting. The authors also acknowledge several limitations, including key compromise and attacks optimized through the public decoder, which have not yet been evaluated.

The new overlap analysis addressed my second concern. The authors have provided additional overlap statistics and controlled experiments. The results indicate that overlap affects regional history recovery mainly through reductions in the remaining size and fragmentation of earlier-generation regions. In particular, smaller and more fragmented regions are more difficult to recover accurately.

Overall, I consider my main concerns to have been sufficiently addressed by the authors' responses and additional experiments. I therefore maintain my original rating of 3.

AuthorsOpenReview ↗

Thank You for Your Thoughtful Review and Constructive Feedback

Dear Reviewer W5b1,

We sincerely appreciate your time, careful consideration, and thoughtful engagement throughout the discussion.

Thank you for thoroughly reviewing our responses and additional experiments, and for acknowledging that our efforts have sufficiently addressed your main concerns. Your constructive feedback on active watermark attacks and regional overlap has been extremely valuable in helping us strengthen both the experimental evidence and the clarity of our explanations.

We also appreciate your observations regarding the remaining limitations. We will ensure that these limitations, along with the effects of region size and fragmentation on history recovery, are clearly discussed in the revised manuscript.

Thank you once again for your careful review, constructive suggestions, and the time you have devoted to evaluating our work. We are truly grateful for your contributions to improving the quality and rigor of our paper.

Best regards, The Authors

Reviewer_qA8COpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper formalizes "reissuable provenance": deciding whether an image is CURRENT, SUPERSEDED, OPEN, or UNKNOWN, and recovering each region's last authorized revision. It proposes an authenticated live head per lineage plus a Generational Stratigraphy Field re-embedded at each reissue.

Critical strengths

The problem is real, and the formalization is clean. The four-way status with explicit abstention (Eq. 2) captures something that single-round embed-edit-detect benchmarks miss: a mark can report the past correctly while the verifier needs the present.

Critical weaknesses

I am not convinced the watermark is doing the work. The live head already stores each issue's scope and parent link (Sec. 3, App. A.2), and Table 3 itself says CURRENT/SUPERSEDED separation is "a contribution of our protocol itself". Sec. 2 even concedes that a C2PA-style system could replay regions along a branch. The obvious baseline, a registry keyed by a robust lineage ID or perceptual hash, is never run. On top of that, the headline numbers come from a friendly setting: OPEN recall of 94.09% (Table 1b) is measured only on four full-regeneration editors that touch almost every pixel, which is where mark destruction is easiest to see. Mark-preserving edits, unseen editors, and adaptive attacks (removal, re-watermarking, forgery) are untested, and the Table 3 baselines were not built to output a regional map. Without these, I can't tell whether GSF adds anything beyond a database lookup.

Questions for authors

How does a registry-only baseline (robust hash or lineage ID plus committed scope and parent link) compare with GSF on status accuracy and regional history?

5 comments

Dear Reviewer qA8C,

Thank you for your careful review and for recognizing that the problem is real and its formalization clean. We are especially grateful for your question about a registry-only design. Although it lies outside the setting our paper studies, it is a valuable additional discussion: answering it let us measure what GSF adds beyond a database lookup.

Our response in brief (Tables R1–R4 are new; the paper's tables are cited by section.)

  1. Clarification: two different settings. Our task verifies from the copy: the verifier has the image, the issuer's key and one live entry per lineage, and regional history "travels inside every circulated copy" (Sec. 1). A registry lookup verifies instead from the issuer's complete per-issue record. The paper distinguishes the two designs (App. M.7), so we take up the lookup as an additional discussion and have now measured it. The robust-ID variant of your proposal, restricted to our task's inputs, is already in the paper (Table 3, Sec. 5.5): it accepts 29.5–38.7% of unissued outputs as issues, against 2.1% for GSF.
  2. Additional discussion: what GSF adds beyond a lookup with no watermark (new experiments). We built the lookup on a content hash and gave it the complete record. Status: on clean copies an exact-hash lookup is exact. Once copies are delivered, one hash distance has to absorb delivery and reject edits, and at each tolerance fixed in advance it fails one of the two; GSF reads additional keyed evidence and, at one operating point frozen before the evaluation, keeps 98.2% of the issues delivered within the range of Sec. 5.4 while accepting 2.2% of the unissued outputs (Table R1). Regional history: a lookup names the issue, not the regions; its map is a replay of the issuer's edit records, exact where they are read and absent where they are not (mIoU 14.3%), whereas GSF reads the map from the copy (79.0%; Table R2). After an unissued local edit, GSF still reads the history of the untouched regions as accurately as before the edit (mIoU 79.4% against 79.7%; Table R3).
  3. Additional experiments: mark-preserving edits, unseen editors and adaptive attacks. Each is tested on all 480 lineages (about 13 GPU-hours in total). Mark-preserving edits: GSF accepts none of 960 copies in which an object is regenerated and pasted back (a robust ID with our protocol: 38.8%), and 3.3–5.8% when an unissued region is pasted into the issue over 5–40% of the field (1,920 copies; the robust ID: 68.5–98.5%). Unseen editors: none of 2,400 outputs is accepted. Adaptive attacks: removal turns none of 2,400 unissued copies into an issue; a third-party watermark over the issue leaves 90.6–98.8% of the issues CURRENT; a forged mark is accepted in at most 3.3% of cases when the blocks that carry the issue evidence are transplanted and 8.3% when the whole mark is copied (global-message watermarks under the same copy attack: 96.5–99.8%; Table R4).

1. Clarification: two different settings

(i) What each design gives the verifier. Our task gives it "a query image, the issuer's key, and a small trusted reference to the lineage's current authorization state", and regional history is read from the copy "without the trusted reference or the issuer's edit records" (Sec. 1). A registry lookup reads more: at every verification, the scope and parent link that the registry records for every issue (Sec. 3) and, for the hash variant, a stored fingerprint of every issue. This is the input that our task does not give the verifier.

(ii) Where GSF's contribution lies. The paper discusses the lookup as an alternative design, with its advantage acknowledged (App. M.7): "That design keeps the history in the registry and reads it online. GSF keeps it in the copy", and such a deployment "needs no field on issued images". GSF's contribution is thus not that a database could not, in principle, provide status and regional history. It is that a verifier without access to the issuer's per-issue edit records can read the history from the copy, recognize unissued edits and obtain issue evidence bound to the image content. We take up your question as an additional discussion and measure what this adds beyond a lookup (part 2).

(iii) What the paper already measures within our task. The robust-ID variant of your proposal, restricted to our task's inputs, is in the paper: rows 1–4 of Table 3 (Sec. 5.5) are robust global-message watermarks that carry the issue ID and run our protocol with the head. They separate CURRENT from SUPERSEDED at 94.7–100%, but accept 29.5–38.7% of unissued outputs (GSF: 2.1%) and carry no inherited history (mIoU 14.4–14.7% at depth 3; GSF: 79.0%). The paper does not run the variant with no watermark; we ran it (part 2).

2. Additional discussion: what GSF adds beyond a database lookup (new experiments)

Setup. A registry without a watermark: for each of the 1,920 issues it stores an exact hash and a 256-bit perceptual hash; a query is matched to the nearest issue of its lineage within t bits (no match: not an issue), and the head then gives CURRENT or SUPERSEDED. We favored the lookup: the lineage is given and the record complete. Three tolerances were fixed in advance, without reference to GSF: t = 0; t = 27, the 99.9th percentile of benign-delivery distances on calibration lineages over the 22 conditions of Table 53 (App. M.6); and t = 12, which loses 2.1% of the delivered current issues over those 22 conditions, which reach beyond the range of Sec. 5.4.

2.1 Status (Sec. 5.2)

Table R1. Status with a registry keyed by a content hash (%). Clean: the queries of Table 1 (Sec. 5.2). Delivered: the nine conditions of Table 53 (App. M.6) within the range of Sec. 5.4.

Registry keyClean issued: CURRENT / SUPERSEDED ↑Clean unissued accepted ↓Delivered current issues kept ↑
Exact hash100.0 / 100.00.00.0
Perceptual hash, t = 0100.0 / 99.70.583.1
Perceptual hash, t = 12100.0 / 99.714.8100.0
Perceptual hash, t = 27100.0 / 99.731.1100.0
Perceptual hash (oracle)100.0 / 99.72.699.4
GSF (reads no per-issue record)99.5 / 99.82.298.2

On clean copies an exact-hash lookup is exact; here we agree with you. Delivery forces a tolerance, and the tolerance admits edits: of the tolerances fixed in advance, those that keep delivered copies accept 14.8% and 31.1% of the unissued outputs, as a robust ID does (29.5–38.7%; Table 3, Sec. 5.5). GSF does not rest on one content distance: it reads keyed evidence (a field in which an edit leaves a disrupted cluster, and issue evidence checked against the issued content; Sec. 4.1) at thresholds frozen before the evaluation (Table 10, App. B.5). Our reply to Reviewer TCKy (Q3) quantifies how moving one of them trades retention against false acceptance under harsher delivery.

2.2 Regional history (Sec. 5.3)

(a) An issued image. A lookup identifies which issue a copy is. Which revision each region comes from is not in a fingerprint; it is in the issuer's edit records.

Table R2. Regional history of the clean current issue I3 (480 lineages; medians, %)

Where the map comes fromMap accmIoU
Lookup alone, no edit record read: every cell gets the identified issue56.314.3
Lookup + replay of every issue's committed scope and parent link (the design of your question)100.0100.0
GSF: read from the copy, with the key and the list of issues93.179.0

The lookup's map is thus a replay of the issuer's records: exact where they are read, as App. M.7 states, and absent where they are not.

(b) A locally edited copy. For an unissued output the paper's target is its parent's field: "an unissued edit does not advance the field" (Sec. 3). A lookup that identifies the parent and replays its records therefore returns the right history; what an ID and a replayed record do not by themselves provide is where the copy departs from its parent. The paper's unissued outputs are complete outputs of an editor (Sec. 5.1): the whole frame is re-synthesized, so no region of such a copy can be assumed to be the parent's. We therefore ran the mark-preserving edit you ask for: in the current issue I3 of each lineage, the largest object (a median of 21.1% of the image) was regenerated with SD-1.5 inpainting and with SDXL-Turbo and pasted back (960 unissued copies). Unlike the paper's full-frame editor outputs, these copies preserve the issue's pixels, and with them its watermark, outside the edited object.

The lookup of Table R1 matches part of these copies to an issue:

Perceptual-hash toleranceUnissued copies matched
t = 00.1%
t = 1211.1%
t = 2742.0%

For the copies it matches, it replays a history and, by the same match, certifies an unissued copy as an issue; for the others it has no history to return. For no copy does it return both the right status and a history.

GSF reads no edit record. It recognizes 99.8% of these copies as unissued, accepts none, and flags 79.3% of the edited cells as disrupted (1.2% of the untouched ones). In the edited region it does not recover the parent's history. On the regions the edit left untouched, it reads the history as accurately as on the issue itself:

Table R3. Regional history after an unissued local edit, read by GSF on the untouched regions (952 copies with untouched cells; medians, %)

Read fromMap accmIoU
The issue I3, before the edit93.679.7
The unissued copy, after the edit93.779.4

The history thus travels with the copy and survives an unissued local edit where the copy is untouched.

3. Additional experiments: mark-preserving edits, unseen editors, adaptive attacks

You are right that the paper's four editors regenerate the whole frame (Sec. 5.1). They are not uniformly destructive, however: a robust ID survives on 29.5–38.7% of their outputs (Table 3, Sec. 5.5), and 18.9 of GSF's 94.0 points of OPEN recall come from field evidence rather than from a destroyed mark (App. M.3). We agree that the three cases you name should be tested. Each experiment below is new, runs on all 480 lineages and uses the paper's verifier unchanged.

Mark-preserving edits. We test two cases. (i) In the 960 locally edited copies of part 2.2, everything outside the edited object is the issue itself; GSF recognizes 99.8% of them as unissued. (ii) In a harder variant, a region of an unissued output is pasted into the current issue over 5%, 10%, 20% or 40% of the field, and the cells that carry the issue evidence are left intact (1,920 copies); GSF recognizes 81.5–90.6% of them as unissued.

Unseen editors. Five editors that the paper does not use (SD-1.5 inpainting returning the whole image, SD-Turbo, LCM, SD-1.5 image-to-image and PeRFlow) each edit the current issue of every lineage (2,400 unissued outputs); GSF recognizes 92.3–100.0% of them as unissued, depending on the editor.

Adaptive attacks. The attacker has the public writer and decoder but not the issuer's key. For removal, VAE re-encoding, diffusion purification at three strengths and bilateral filtering are each applied to the paper's unissued output of the current issue in every lineage (2,400 copies). For re-watermarking, a third-party watermark (TrustMark-Q, VINE-R or WAM) is embedded over the current issue. For forgery, the target is one of four kinds of unissued copy (an editor's complete output of the current issue or of an earlier issue; the issue with 40% of its cells replaced; the issue with one object regenerated). The attacker either transplants onto it the pixel blocks of the current issue that carry the issue evidence, or decodes the issue's whole mark with the public decoder and re-embeds it in the target with the public writer (1,920 copies for each forgery).

Table R4. Unissued copies accepted as issues in the new experiments (%; ranges are over the four pasted areas or the four kinds of unissued copy)

ExperimentUnissued copiesGSF ↓For reference ↓
Local edit pasted back9600.0robust ID (TrustMark-Q) + our protocol, same edits: 38.8
Unissued region pasted into the issue1,9203.3–5.8the same robust ID, same construction: 68.5–98.5
Unseen editors2,4000.0–
Removal2,4000.0–
Forgery: transplant of the evidence blocks1,9200.0–3.3–
Forgery: copy of the whole mark1,9200.4–8.3global-message watermarks, same attack: 96.5–99.8

GSF's issue evidence is tied to the issued content (Sec. 4.1), so evidence moved onto other content rarely matches. A global message is not checked against the content: the same copy attack, run on the editor's output of the current issue, is accepted in 96.5–99.8% of cases for VINE-R, TrustMark-Q and OmniGuard, where GSF accepts 7.5% on that kind of copy. On the issue itself, removal takes the status away (at most 0.2% remain CURRENT), whereas a third-party watermark leaves 90.6–98.8% of the issues CURRENT. On baselines that output a regional map, please see our reply to Reviewer TCKy (Q1): a full-field-refresh baseline built from WAM stays 30 / 44 / 45 mIoU points below GSF at depths 1–3, at matched image quality.

In closing. GSF's contribution is not that a database could not, in principle, provide status and regional history. It is that a verifier without access to the issuer's per-issue edit records can read the regional history from the copy (Table R2), including on the untouched regions of an unissued local edit (Table R3); recognize unissued edits (Table 3, Sec. 5.5, and Table R1); and obtain issue evidence bound to the image content (Table R4). A system with similar functions could be built on a database that holds the issuer's complete record, by identifying the parent issue, retrieving its recorded history and comparing image content; but that takes additional inputs, a content-comparison step and deployment assumptions, and is no longer a plain ID lookup. GSF embeds this evidence in the image, so that it travels with the copy. We will add this discussion and the new experiments to the appendix; the claims of the paper are unchanged. We hope this addresses your concern, and we would be grateful if you would reconsider your assessment.

Sincerely, The Authors

Comments on the paper

3 comments
Area_Chair_URYZOpenReview ↗

Participate in the discussion period

Dear reviewers,

The authors have replied to your comments. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!

Best, Your AC

AuthorsOpenReview ↗

General response

Dear Reviewers and Area Chair,

We sincerely thank the three reviewers for their careful and insightful comments, and are pleased that all three recognize the importance of the problem and the contribution of our formalization.

The reviews' questions have a common form: each asks for evidence or clarification that the submission did not contain. Our replies now provide it, from experiments on the paper's 480 lineages with GSF as submitted (apart from the threshold sweep for Reviewer TCKy's Q3); the requested registry lookup and regional baseline are set up in their favor. The results, Tables R1–R9, change no conclusion of the paper.

Reviewer qA8C raised an important question, "whether GSF adds anything beyond a database lookup", and helpfully identified five evaluations that would settle it: a registry lookup, mark-preserving edits, unseen editors, adaptive attacks, and a baseline that outputs a regional map. We have now measured all five (the last under Reviewer TCKy's Q1).

  • The registry lookup (Tables R1–R3) reads the issuer's complete per-issue record, which our task does not give the verifier (an alternative design, App. M.7). Built on a content hash and given that record, it accepts 14.8% and 31.1% of the unissued outputs at the tolerances, fixed in advance, that keep delivered copies; GSF accepts 2.2% and keeps 98.2% of the issues delivered within the range of Sec. 5.4. The lookup's map is exact when the issuer's edit records are read and reaches 14.3% mIoU when they are not; GSF reads 79.0% from the copy and, after an unissued local edit, still reads the untouched regions (79.4% against 79.7% before the edit).
  • Mark-preserving edits, unseen editors, adaptive attacks (Table R4). GSF accepts none of 960 copies with an object regenerated and pasted back, 3.3–5.8% of 1,920 copies with an unissued region pasted over 5–40% of the field, and none of 2,400 outputs of five unseen editors. Removal turns none of 2,400 unissued copies into an issue; a third-party watermark leaves 90.6–98.8% of the issues CURRENT; a copied mark is accepted in at most 8.3% (global-message watermarks, same attack: 96.5–99.8%).

Reviewer TCKy raised three valuable questions, which help clarify both the strengths and the boundaries of our approach. Each now has a dedicated experiment.

  • Q1, regional baseline (Table R5). A full-field-refresh WAM baseline, tuned in its favor, stays 30.3 / 43.7 / 44.8 mIoU points below GSF on inherited history (depths 1–3) at matched image quality. With full registry-history access, the map of an issued image is a lookup for any carrier (App. M.7): the lookup above shows what GSF adds.
  • Q2, longer histories (Table R6). From four to eight versions, Map acc / mIoU go from 93.1 / 79.0% to 86.9 / 59.0%; cells without a unique anchor are still read at 84.8%. The loss sits in small fragments; beyond the declared depth a three-bit code can do better, as Sec. 4.2 states.
  • Q3, harsh delivery (Table R7). Sweeping one threshold raises retention under JPEG 50 / Platform B from 35.0 / 2.5% to 93.8 / 90.4%, and the unissued edits accepted from 0.5 / 0.0% to 3.4 / 2.1% there and from 2.5% to 8.4% on clean copies; TrustMark-Q retains 98.3 / 99.2% and accepts 24.0 / 26.1%. The intended envelope is the 18 conditions of App. E.3; these two lie outside it.

Reviewer W5b1 offered thoughtful suggestions on the scope, limitations and evaluation of active attacks (Q1) and on overlap statistics, sensitivity and overlap levels (Q2). We have now addressed all six points.

  • Q1, active attacks (Table R8). The reply states the scope (the submission tests no attack designed against the mark), runs removal, disruption and forgery by an attacker without the key, and lists four limitations. An issue can be made unverifiable (after removal, 1 of 2,400 is still CURRENT); unissued copies are accepted at the rates of Table R4.
  • Q2, overlap (Table R9). The approved regions' median IoU is 0.43 for non-adjacent and 0.45 / 0.44 for adjacent issues. Across quartiles of observed overlap, Map acc varies by at most 0.6 points and mIoU falls by 4.0; under controlled overlap they fall by up to 2.3 and 15.6 points.

Taken together, a verifier without access to the issuer's per-issue edit records can read the regional history from the copy, recognize unissued edits and obtain issue evidence bound to the image content. Where the mark is removed or delivery is harsher than the intended envelope, the cost falls on retention: the issue reads OPEN or UNKNOWN. The revision will add the stronger baseline to Table 3 (Sec. 5.5), scope and limitations to Sec. 5.4, and the other experiments to the appendix.

We hope that these experiments and clarifications have addressed the reviewers' concerns. We would greatly appreciate any further feedback, on points that remain unresolved or on new questions, during the discussion period.

Thank you again for your time and consideration.

Sincerely, The Authors

AuthorsOpenReview ↗

Summary of the Discussion

Dear Area Chair and Reviewers,

We thank the Area Chair and the three reviewers for a careful and constructive discussion. In brief: every question in the reviews now has a dedicated experiment or analysis, the additional results support the paper's conclusions, and two of the three reviewers have confirmed that their concerns are addressed.

The paper introduces and formalizes reissuable provenance: determining whether an image is the current issue, a superseded issue or an unissued edit, with abstention when evidence is insufficient, and recovering each region's last authorized revision. GSF combines watermark evidence with an authenticated live reference and rewrites its field in full at each reissue, so history is preserved without requiring old watermarks to survive successive edits. The evaluation follows 480 histories through four releases and retains each AI editor's entire output. All three reviews recognize the problem and its formalization: "The problem is real, and the formalization is clean" (qA8C); a "practically relevant problem and a useful step beyond existing approaches" (W5b1); "a coherent problem formulation and reissuance protocol" (TCKy).

Where the discussion stands. The new experiments use the paper's 480 lineages, with GSF's writer, decoder and reader as submitted.

ReviewerQuestionsEvidenceFeedback (rating)
qA8Cregistry lookup; mark-preserving edits; unseen editors; adaptive attacks; a baseline that outputs a regional mapTables R1–R5, R12not yet received (2)
TCKystronger regional baseline; longer histories; harsh deliveryTables R5–R7"address my questions, and I have raised my score to 3" (2 → 3)
W5b1active attacks; overlap among approved regionsTables R8–R9"I consider my main concerns to have been sufficiently addressed" (3, maintained)

Reviewer qA8C raised the central question of the discussion. The review lists five evaluations as missing and concludes: "Without these, I can't tell whether GSF adds anything beyond a database lookup." All five are now measured. A lookup gives the verifier more than the paper's task does, so we answer in both settings.

  • The setting of the paper. The verifier has the image, the issuer's key and one live entry per lineage, and regional history "travels inside every circulated copy" (Sec. 1). The robust-ID variant of the proposed registry, restricted to these inputs, is already in the paper (Table 3, Sec. 5.5): robust global-message watermarks that carry the issue ID and run our protocol accept 29.5–38.7% of unissued outputs as issues (GSF: 2.1%) and carry no inherited history.
  • The setting of the question. A registry lookup verifies from the issuer's complete per-issue record, an alternative design that the paper distinguishes (App. M.7). As an additional discussion we built it on a content hash, without a watermark, and gave it the complete record. At the hash tolerances, fixed in advance, that keep delivered copies, it accepts 14.8% and 31.1% of the unissued outputs; with GSF alongside, 1.9% and 2.1% (Table R12). The record supplies the identity and the exact history; the field recognizes most of the unissued edits that the tolerance admits and shows where a copy was changed.

GSF thus adds evidence beyond a database lookup in both deployments. On the other four requests: GSF accepts as issues few or none of the mark-preserving edits and none of 2,400 outputs of five unseen editors (Table R4); the attacks and the regional baseline, also raised by Reviewers W5b1 and TCKy, follow below.

Reviewer TCKy asked for a stronger regional baseline, longer histories and harsher delivery. A full-field-refresh baseline, tuned in its favor, stays well below GSF on inherited history. The three experiments delimit where the method is most advantageous, and, as the reviewer suggested, the Conclusion now states these settings: verification without access to per-issue records, histories of about four issues, and the delivery envelope of App. E.3.

Reviewer W5b1 asked about active attacks and overlap among approved regions. An attacker without the key can make an issue unverifiable; with a copied mark, at most 8.3% of unissued copies pass as issues (global-message watermarks, same attack: 96.5–99.8%). Overlap acts through the size of what remains of earlier regions. The threat model and the limitations are stated in the new App. P.

The revision (uploaded; changes in blue; tables numbered as in our replies) adds the stronger baseline to Table 3 and the new experiments to App. E.6, G.1–G.2, G.5, L.10, M.8–M.9, O and P.

The task, the reissuance protocol and the complete-output evaluation stand as submitted. We thank Reviewers TCKy and W5b1 for confirming that their concerns are addressed. We have not yet heard from Reviewer qA8C and would be grateful if the Area Chair could consider the new results alongside the review.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.