Image Watermarking With Memory Across Reissuance
Organizations: University of Southern California
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Publishers often revise and reissue images, yet valid watermarks may persist on superseded releases or unapproved edits.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 We introduce and formalize reissuable provenance: determining an image's current, superseded, or unissued status and recovering each region's last authorized revision (its generation), with abstention when evidence is insufficient.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 GSF achieves 99.67% accuracy in classifying issued images as current or superseded and 94.09% recall of unissued edits.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.
Figures & tables
| True status | Overall | ||||
|---|---|---|---|---|---|
| Current | 99.58 | 99.79 | 99.79 | 98.96 | 99.53 |
| Superseded | – | 99.58 | 99.79 | 99.79 | 99.76 |
| Ambiguous cells | Whole field and newest slice | ||||||
| Depth | Ambig. | Cell acc. | Spatial acc. | Map acc. | Map mIoU | Newest IoU | Newest F1 |
| 1 | 24.38 | 46.67 | 88.64 | 97.19 | 93.87 | 94.92 | 97.39 |
| 2 | 43.75 | 35.14 | 89.23 | 95.00 | 87.27 | 94.09 | 96.95 |
| 3 | 58.13 | 25.56 | 88.27 | 93.13 | 78.96 | 92.86 | 96.30 |
| Issued images | Unissued outputs | Generation-map mIoU | ||||||
|---|---|---|---|---|---|---|---|---|
| Existing carrier our reissuance protocol | Current | Superseded | Open | False issue | Unknown | |||
| VINE-R | 98.2 | 98.9 | 63.5 | 36.5 | – | 28.91 | 19.40 | 14.70 |
| TrustMark-Q | 99.9 | 100.0 | 70.1 | 29.9 | – | 28.52 | 19.27 | 14.40 |
| TrustMark-RM | 99.6 | 99.8 | 70.5 | 29.5 | – | 28.52 | 19.21 | 14.65 |
| OmniGuard tamper mask | 95.3 | 94.7 | 61.3 | 38.7 | – | 28.32 | 19.08 | 14.65 |
| WAM in GSF layout † | 20.4 | 40.8 | 50.8 | 0.4 | 48.8 | 49.05 | 32.38 | 24.30 |
Appendix figures & tables68 assets
Supplementary material from the paper’s appendix.
Appendix
| Query event | Event-defined truth | Supported issue evidence | Head role | Regional target |
|---|---|---|---|---|
| Issue | Current | matches | selects as live | |
| Issue , | Superseded | matches | establishes | |
| Any issue, | Superseded | matches its issue | declares no issue active | |
| Unissued output from | Open | no committed issue match | none until issuance | or |
| Insufficient/out-of-lineage evidence | Unknown | unsupported | cannot be applied |
| Artifact | Role |
|---|---|
| Writer–decoder checkpoint | physical two-bit channel |
| Carrier/verification core | layout, coding, synchronization |
| Protocol implementation | issue evidence and trajectory execution |
| Spatial-reader checkpoint | candidate-equivariant recovery |
| Trajectory assignment | roots, prompts, parents, seeds |
| Evaluation image set | roots and derived lifecycle images |
| Component | Cells | Physical slots | Coding |
|---|---|---|---|
| Lineage evidence | 32 | 64 | 36 bits BCH |
| Issue evidence | 64 | 128 | 64 bits BCH |
| Generation field | 160 | 320 | two keyed bits per cell |
| Total | 256 | 512 | shared writer/decoder |
| Stage | Operation | Channels | Spatial size |
|---|---|---|---|
| Input | RGB two-bit map | 5 | |
| Stem | , stride 1 | ||
| Down 1 | , stride 2 | ||
| Down 2 | , stride 2 | ||
| Middle | two blocks | ||
| Up 2 | nearest upsample, skip, |
| Stage | Operation | Channels | Spatial size |
|---|---|---|---|
| Stem | , stride 1 | ||
| Down 1–2 | two , stride 2 | ||
| Down 3–5 | three , stride 2 | ||
| Context | two blocks | ||
| Readout | convolution |
| Check | Observation | Protocol significance |
|---|---|---|
| Clean output | PSNR : 45.646 dB; LPIPS : 0.003431; median maximum error : 13 | Establishes the physical fidelity of the frozen carrier; the fidelity distribution of the 480 evaluation roots appears in Appendix J . |
| JPEG 75, resize to 384 and back, rotate | BER : 0.00098, 0.00000, and 0.00081 | Confirms payload readout under representative delivery operators before status routing. |
| Screenshot simulation | BER : 0.01571 | Measures the harder composite delivery channel used in the training curriculum. |
| Crop 90% and resize | all-grid BER : 0.07723 | Exposes missing border support; the verifier accounts for that support before interpreting the payload. |
| Evidence family | Frozen rule | Placement rule (development data) | Protocol role |
|---|---|---|---|
| Cell readability | readable iff margin , | Maximum of the per-root first-percentile margins over 12 development roots and five delivery transforms after predictable erasures (14,304 cell observations). | Produces the erasure mask consumed by downstream status inference. |
| Damage routing on clean queries | , , | The three coordinates were fixed jointly at one operating point on 144 clean queries from 12 development roots and 20 additional roots selected by a fixed hash rule. The status objective maximized event-defined Open while preserving status correctness on issued images and bounding false Current and issued-image Open outcomes. | Separates localized disruption supporting Open from complete loss routed to Unknown . |
| Unmatched field candidates | Smallest integer covering the benign calibration maximum while keeping the row-level false-trigger bound below ; confirmed on a separate 12-root calibration split. | More than one live cell matching no issued candidate supplies positive Open evidence. | |
| Issue consistency | 99th percentile over 359 benign queries from 60 development roots, rounded up to the native distance grid. | Excess mismatch supplies positive Open evidence. | |
| Current-issue safeguard | within , | The pose predicate was fixed before the evaluation; the mismatch threshold is the rounded-up 99th percentile of 240 benign identity-pose queries on the same 60 roots. | Withholds Current when current-issue evidence is insufficient, routing the internal rejection to Unknown . |
| Object | Count |
|---|---|
| COCO-derived roots | 480 |
| Issues per root | 4 |
| Complete-output edits per root | 5 |
| Issue operations | |
| Complete-output editor calls | 2,400 |
| Stored images per root | 10 |
| Editor | Precision | Inference configuration | Edges |
|---|---|---|---|
| FLUX.1-Kontext-dev | bfloat16 | 28 steps, guidance 2.5 | 592 |
| InstructPix2Pix | float16 | 30 steps, text guidance 8.0, image guidance 1.5 | 593 |
| Qwen-Image-Edit-2509 | bfloat16 | fused Lightning LoRA, 8 steps, CFG 1.0 | 591 |
| SDXL-Turbo img2img | float16 | 8 steps, guidance 0.0, strength 0.50 | 624 |
| Mean absolute RGB change | Changed-cell fraction | |||||
|---|---|---|---|---|---|---|
| Editor | p10 | median | p90 | minimum | p10 | median |
| FLUX.1 Kontext | 4.95 | 16.18 | 81.06 | 0.09 | 0.80 | 1.00 |
| InstructPix2Pix | 13.82 | 34.77 | 70.59 | 0.60 | 1.00 | 1.00 |
| Qwen Image Edit | 5.28 | 17.26 | 51.23 | 0.18 | 0.74 | 1.00 |
| SDXL-Turbo | 24.75 | 32.55 | 41.64 | 1.00 | 1.00 | 1.00 |
| Ground-truth class | Query condition | Query set |
|---|---|---|
| Current | each issue under its own live head | |
| Superseded | each predecessor under every later head | |
| Open | pre-commit outputs later closed on the main chain | |
| Open | unissued output from the terminal issue | |
| Open | derivative from issue 1, queried at head 1 | 249 |
| Open | derivative from issue 1, queried at head 2 | 249 |
| Ground truth | Query condition | Correct | Open | Unk. | Accuracy (%) |
| Current | issue 0 at head 0 | 478 | 2 | 0 | 99.58 |
| issue 1 at head 1 | 479 | 1 | 0 | 99.79 | |
| issue 2 at head 2 | 479 | 1 | 0 | 99.79 | |
| issue 3 at head 3 | 475 | 2 | 3 | 98.96 | |
| all conditions under the issue’s own head ( ) | 1911 | 6 | 3 | 99.53 | |
| Superseded | issue 0 at head 1 | 478 | 2 | 0 | 99.58 |
| Query condition | Query set | Open | False issue | Unk. | Open recall (%) |
| pre-issue-1 output at head 0 | 480 | 461 | 9 | 10 | 96.04 |
| pre-issue-2 output at head 1 | 480 | 452 | 11 | 17 | 94.17 |
| pre-issue-3 output at head 2 | 480 | 446 | 10 | 24 | 92.92 |
| pre-reissue editor outputs | 1359 | 30 | 51 | 94.38 | |
| final unissued output at head 3 | 440 | 12 | 28 | 91.67 | |
| derivative from issue 1 at head 1 | 249 | 234 | 6 | 9 | 93.98 |
| Routing rule | Issued acc. | Open recall | Issue FAR |
|---|---|---|---|
| Abstain on broad field loss | 99.88 | 14.29 | 4.92 |
| Abstain only on complete field loss | 99.88 | 88.65 | 4.92 |
| read any disrupted cluster as Open | 99.67 | 94.09 | 2.24 |
| Oracle upper bound (subset) | – | 98.23 | – |
| Final prediction | Dominant evidence route | Count | Fraction (%) |
| Open | issue evidence does not decode | 2345 | 74.94 |
| localized carrier-disruption support | 543 | 17.35 | |
| additional inconsistency evidence | 56 | 1.79 | |
| Unknown | insufficient or broadly degraded evidence | 115 | 3.68 |
| false issue | false Current match | 49 | 1.57 |
| false Superseded match | 21 | 0.67 |
| Current issue | Initial issue | Open | Pooled and | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Condition | Current | Open | Unk. | Superseded | Open | Unk. | Open | False issue | Unk. | ||
| Clean | 98.96 | 0.42 | 0.63 | 99.79 | 0.21 | 0.00 | 91.67 | 94.58 | 93.13 | 2.50 | 4.38 |
| JPEG 95 | 99.58 | 0.00 | 0.42 | 99.79 | 0.21 | 0.00 | 92.50 | 96.46 | 94.48 | 2.40 | 3.13 |
| JPEG 90 | 98.96 | 0.42 | 0.63 | 99.38 | 0.63 | 0.00 | 93.13 | 96.25 | 94.69 | 2.08 | 3.23 |
| JPEG 85 | 97.92 | 1.25 | 0.83 | 99.17 | 0.83 | 0.00 | 94.38 | 96.46 | 95.42 | 1.88 | 2.71 |
| JPEG 75 | 95.63 | 3.96 | 0.42 | 96.46 | 3.54 | 0.00 | 95.00 | 97.92 | 96.46 | 1.77 | 1.77 |
| Map acc. | Map mIoU | Newest F1 | Current at own head | |||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Condition | ||||||||||
| Clean | 97.19 | 95.00 | 93.13 | 93.87 | 87.27 | 78.96 | 96.30 | 99.79 | 99.79 | 98.96 |
| JPEG 95 | 97.50 | 95.00 | 93.13 | 93.84 | 87.11 | 79.18 | 96.30 | 99.58 | 99.79 | 99.58 |
| JPEG 90 | 96.88 | 95.00 | 93.13 | 93.87 | 87.11 | 79.12 | 96.37 | 99.58 | 99.38 | 98.96 |
| JPEG 85 | 96.88 | 95.00 | 93.13 | 93.85 | 87.00 | 79.03 | 96.34 | 98.75 | 97.92 | 97.92 |
| JPEG 75 | 96.88 | 95.00 | 93.13 | 93.81 | 86.96 | 79.05 | 96.32 | 96.25 | 96.67 | 95.63 |
| Clipped pixels (%) at brightness | ||||||
|---|---|---|---|---|---|---|
| Image | Mean | 99th pct. | ||||
| (initial issue) | 109.06 | 241.53 | 3.83 | 5.79 | 7.98 | 14.62 |
| (issue 1) | 107.56 | 231.72 | 2.65 | 4.53 | 6.64 | 13.22 |
| (issue 2) | 108.99 | 227.60 | 1.99 | 3.92 | 5.95 | 12.58 |
| (current issue) | 106.94 | 221.84 | 1.64 | 3.12 | 4.86 | 11.26 |
| (output after ) | 108.50 | 221.92 | 1.68 | 3.36 | 5.24 | 12.03 |
| JPEG 50 | Platform B | Clean | |
|---|---|---|---|
| 0 | 1.9 / 0.0 | 0.0 / 0.0 | 98.1 / 1.0 |
| 1 (Table 10 ) | 35.0 / 0.5 | 2.5 / 0.0 | 99.0 / 2.5 |
| 2 | 54.0 / 0.8 | 11.7 / 0.1 | 99.4 / 3.1 |
| 3 | 68.5 / 1.0 | 29.8 / 0.4 | 99.4 / 3.8 |
| 4 | 75.2 / 1.6 | 46.9 / 0.5 | 99.4 / 4.6 |
| 6 | 84.8 / 2.2 | 68.3 / 0.8 | 99.4 / 6.1 |
| Ambiguity | Ambiguous-cell accuracy | |||||
|---|---|---|---|---|---|---|
| Depth | Theory (%) | Observed (%) | Cells | Local | Spatial | Gain |
| 1 | 25.00 | 24.38 | 39 | 46.67 | 88.64 | +41.97 |
| 2 | 43.75 | 43.75 | 70 | 35.14 | 89.23 | +54.09 |
| 3 | 57.81 | 58.13 | 93 | 25.56 | 88.27 | +62.71 |
| Inference operator | Depth-1 acc. | Depth-2 acc. | Depth-3 acc. |
|---|---|---|---|
| Independent local likelihood | 46.67 | 35.14 | 25.56 |
| Hard partition + exact region decoding | 84.21 | 82.02 | 80.23 |
| Amortized soft spatial reader | 88.64 | 89.23 | 88.27 |
| Recurrent soft spatial reader | 88.89 | 88.66 | 89.01 |
| Mechanism | Ambiguous-cell accuracy (%) | Increment |
|---|---|---|
| Independent local likelihood | 36.97 | – |
| Learned evidence recalibration, no spatial kernel | 67.48 | +30.51 |
| soft multiscale spatial kernels | 87.83 | +20.35 |
| Recurrent soft coupling (three seeds) | 87.70–88.11 | |
| Hard partition + exact decoding | 80.78 |
| Label | Present | Minimum | Q1 | Median | Q3 | Maximum |
|---|---|---|---|---|---|---|
| 480 | 1 | 18 | 33 | 52 | 168 | |
| 466 | 0 | 9 | 19 | 33 | 102 | |
| 476 | 0 | 25 | 44 | 71 | 153 | |
| 480 | 64 | 126 | 152 | 170 | 220 |
| Independent local likelihood | Spatial generation-field reader | |||||||
| Depth | Map acc. | Map mIoU | Newest IoU | Newest F1 | Map acc. | Map mIoU | Newest IoU | Newest F1 |
| 1 | 86.88 | 75.82 | 75.00 | 85.71 | 97.19 | 93.87 | 94.92 | 97.39 |
| 2 | 71.88 | 52.61 | 56.09 | 71.87 | 95.00 | 87.27 | 94.09 | 96.95 |
| 3 | 57.50 | 36.76 | 42.19 | 59.34 | 93.13 | 78.96 | 92.86 | 96.30 |
| Presence-conditioned IoU (%) | ||||
| Depth | ||||
| 1 | 93.99 | 94.92 | – | – |
| 2 | 89.32 | 83.33 | 94.09 | – |
| 3 | 81.48 | 72.57 | 81.32 | 92.86 |
| Share of the previous band covered | 0% | 25% | 50% | 75% | 100% |
|---|---|---|---|---|---|
| Lines still carrying and (of 4) | 4 | 3 | 2 | 1 | 0 |
| Map acc. (mean) | 98.2 | 97.5 | 96.9 | 95.9 | 97.8 |
| mIoU (mean) | 96.5 | 94.4 | 91.2 | 80.9 | 95.8 |
| Newest slice read correctly | 98.4 | 98.4 | 98.5 | 98.6 | 99.1 |
| and read correctly | 98.4 | 96.7 | 93.4 | 80.8 | – |
| Depth | Predicted Current (%) | Ambiguous-cell acc. (%) | Map accuracy (%) | Newest IoU (%) |
|---|---|---|---|---|
| 1 | 99.79 | 88.64 | 97.19 | 94.92 |
| 2 | 99.79 | 89.23 | 95.00 | 94.09 |
| 3 | 98.96 | 88.27 | 93.13 | 92.86 |
| Depth | 3 (Sec. 5.3 ) | 4 | 5 | 6 | 7 |
|---|---|---|---|---|---|
| Map acc. | 93.1 | 91.3 | 90.0 | 88.1 | 86.9 |
| mIoU | 79.0 | 72.5 | 67.1 | 62.8 | 59.0 |
| Ambiguous cells, observed / Eq. ( 5 ) | 58.1 / 57.8 | 68.8 / 68.4 | 76.3 / 76.3 | 82.5 / 82.2 | 86.9 / 86.7 |
| Ambiguous-cell accuracy | 88.3 | 87.4 | 86.9 | 85.7 | 84.8 |
| Three-bit code, error-free upper bound: map acc. / mIoU | 90.6 / 76.0 | 90.3 / 73.8 | 90.3 / 72.1 | 89.9 / 70.4 | 89.9 / 69.4 |
| Intermediate canvas | Kernel pairs | Carrier score | Clean ratio | Current |
|---|---|---|---|---|
| None (clean) | 1 | 17.999 | 1.000 | |
| 768 | 2 | 18.013–18.042 | 1.002–1.003 | |
| 1024 | 6 | 17.929–18.040 | 0.996–1.002 | |
| 1536 | 2 | 18.016–18.030 | 1.001 | |
| 2048 | 2 | 18.030–18.048 | 1.001–1.002 |
| Condition | Native 512 | Canonical hub | Difference |
|---|---|---|---|
| Clean | 1.0000 | 1.0000 | 0.0000 |
| JPEG 75 | 1.0000 | 1.0000 | 0.0000 |
| Downsample/upsample | 1.0000 | 1.0000 | 0.0000 |
| Crop 90% and resize | 0.8000 | 0.9333 | |
| Rotate | 1.0000 | 1.0000 | 0.0000 |
| Screenshot simulation | 0.9667 | 1.0000 |
| Method | Released weight | Localization signal | Native threshold |
|---|---|---|---|
| EditGuard | clean | revealed blue residual | 0.2 |
| EditGuard | degradation-trained | revealed blue residual | 0.4 |
| OmniGuard | DTE checkpoint 1500 | sigmoid tamper probability | 0.5 |
| Native paste-back | Retained complete output | |||
|---|---|---|---|---|
| Method | F1 | Pred. area | F1 | Pred. area |
| EditGuard, clean weight | 0.9718 | 0.0962 | 0.1681 | 1.0000 |
| EditGuard, degradation weight | 0.9519 | 0.0850 | 0.1681 | 1.0000 |
| OmniGuard | 0.9847 | 0.0935 | 0.0000 | 0.0027 |
| Diagnostic | Clean | Paste-back | Complete output |
|---|---|---|---|
| EditGuard clean-weight message BER | 0.0000 | 0.0000 | 0.5000 |
| EditGuard degradation-weight message BER | 0.0000 | 0.0000 | 0.5156 |
| OmniGuard artifact PSNR (dB) | 38.66 | 17.35 | 6.67 |
| GSF disrupted-cell fraction | 0.0000 | 0.0837 | 0.8778 |
| Method | Carrier diagnostic | Pred. area | F1 | Output form |
|---|---|---|---|---|
| EditGuard, clean weight | BER : 0.5391 | 1.0000 | 0.7495 | all positive |
| EditGuard, degradation weight | BER : 0.4844 | 1.0000 | 0.7532 | all positive |
| OmniGuard | artifact PSNR : 6.57 dB | 0.0039 | 0.0096 | nearly empty |
| PSNR (dB) | ||||||
|---|---|---|---|---|---|---|
| Minimum | P10 | Mean | Median | P90 | Maximum | |
| 480 | 36.850 | 42.504 | 45.973 | 46.694 | 48.141 | 48.722 |
| Result | Unit | Denominator | Aggregation and dependence |
|---|---|---|---|
| Status of issued images | root–head query | own-head and later-head queries | Pooled fixed-condition accuracy; roots cluster repeated heads |
| Status of unissued outputs | root–head query | pre-reissue, final-unissued, and four earlier-issue conditions ( ) | Pooled recall; event truth fixed before evidence; roots cluster repeated queries |
| Post-processing robustness | root–condition query | The clean images and 31 prespecified transformations on reused roots | Per-condition rates followed by median/minimum stress summaries; conditions are not iid draws |
| Collision recovery | successor root | 480 successors independently at each tested depth | Ambiguous-cell accuracy is computed within image, then summarized across roots |
| Generation maps | successor root | 480 successors at each tested depth | Per-image accuracy/IoU, then median; absent labels do not create artificial zero-IoU classes |
| Matched endpoint control | image | 1,000 valAGE images | Paired by source, mask, editor configuration, and seed within each carrier |
| Component | Evaluated cost | Scaling or role |
| Fixed distributed representation | ||
| Writer | 596,355 parameters | one full rewrite per issuance |
| Decoder | 2,329,155 parameters | one canonical evidence field |
| Spatial reader | 4,633 parameters | shared over candidate and spatial axes |
| Learned total | 2,930,143 parameters | independent of lineage depth |
| Image-resident payload | 512 raw binary slots | 2 slots per cell; depth independent |
| Axis | Evidence | Established result |
|---|---|---|
| Status semantics | Current/superseded/open behavior across repeated complete-output editing, branching, and head changes on 480 roots | A live head expires old issues, while reissuance separates unissued edits from issued successors |
| Regional history | Four coexisting labels, depth-conditioned collisions, and generation-map recovery | One constant-rate field recovers temporal attribution and current revision support |
| Resolution | Canonicalized delivery from 768 through 2048 pixels | Resolution changes preserve keyed cell identity and all tested issues |
| Post-processing | Paired issued/unissued queries under 31 frozen transformations | Routine delivery preserves both status decisions; heavier stress reveals their distinct evidence curves |
| Authorized scope | Issuer commitment of a blockwise pixel-difference proposal derived from each complete editor return | The newest generation slice localizes where the current revision entered authorized history |
| External controls | Matched native-paste-back versus retained-output endpoint under a shared SD-1.5 inpainting setup | Complete outputs remove the predecessor-residual premise used by snapshot localizers |
| Property | Direct two-bit tag | GSF keyed field |
|---|---|---|
| Intact labels – | Collision-free bijection | Predictable local stochastic ambiguity |
| Candidate growth | and alias at every cell | Same interface; unique anchors decrease with candidate count |
| Encoded relation | Generation modulo four | Compatibility for |
| Damaged symbol | Every two-bit value remains a legal label | Exact matching can return no match or multiple matches |
| Spatial evidence | A global alias leaves no distinguishing position | A pair may collide here and differ elsewhere |
| Wider code | Postpones alias but reduces fixed-slot spatial resolution | Retains the two-bit, 160-cell operating point |
| Local outcome | Direct lookup | Keyed exact |
|---|---|---|
| Correct generation | 42.04% | 17.60% |
| Wrong generation | 57.96% | 27.52% |
| Abstain / ambiguous | 0.00% | 54.88% |
| Generation accuracy | ||||
|---|---|---|---|---|
| Query | Cells/image | Keyed exact | Direct lookup | GSF spatial |
| at head 3 | 31.8 | 0.0000 | 0.0000 | 0.2248 |
| at head 3 | 32.7 | 0.0000 | 0.0000 | 0.2188 |
| at head 3 | 31.0 | 0.0000 | 0.0000 | 0.2355 |
| at head 3 | 31.4 | 0.0000 | 0.0000 | 0.2303 |
| at head 3 | 31.6 | 0.0000 | 0.0000 | 0.2377 |
| Condition | Direct agreement | Permuted-direct agreement | Sign asymmetry |
|---|---|---|---|
| Clean | 0.8499 | 0.8471 | 0.0058 |
| Blur | 0.7677 | 0.8191 | 0.0037 |
| Brightness | 0.8567 | 0.8535 | 0.0047 |
| Brightness | 0.8544 | 0.8271 | 0.0069 |
| Contrast | 0.8673 | 0.8465 | 0.0088 |
| JPEG 50 | 0.7247 | 0.7711 | 0.0014 |
| Expected unique anchors | |||||
|---|---|---|---|---|---|
| 1 | 160.00 | 120.00 | 92.75 | 75.00 | 39.84 |
| 2 | 80.00 | 90.00 | 81.16 | 70.31 | 39.69 |
| 3 | 40.00 | 67.50 | 71.01 | 65.92 | 39.53 |
| 4 | 20.00 | 50.62 | 62.14 | 61.80 | 39.38 |
| 5 | 10.00 | 37.97 | 54.37 | 57.94 | 39.22 |
| Added bit | 2nd | 3rd | 4th | 5th | 6th | 7th | 8th |
|---|---|---|---|---|---|---|---|
| Break-even depth | 2 | 3 | 5 | 7 | 12 | 20 | 34 |
| Theory | Observed | Theory | Observed | Theory | Observed | |
|---|---|---|---|---|---|---|
| 1 | 0.5000 | 0.4997 | 0.2500 | 0.2514 | 0.1250 | 0.1267 |
| 2 | 0.7500 | 0.7513 | 0.5625 | 0.5637 | 0.4219 | 0.4234 |
| 3 | 0.8750 | 0.8763 | 0.7656 | 0.7670 | 0.6699 | 0.6728 |
| 4 | 0.9375 | 0.9365 | 0.8789 | 0.8775 | 0.8240 | 0.8224 |
| 8 | 0.9961 | 0.9962 | 0.9922 | 0.9918 | 0.9883 | 0.9892 |
| Metric | Keyed 2-bit | Keyed 8-bit | Eight-bit change |
| Conditional reader accuracy | |||
| All valid generation slots, accuracy | 0.8539 | 0.8599 | |
| Decoded-code-flip slots, accuracy | 0.2743 | 0.3031 | |
| Carrier-destroyed slots, accuracy | 0.4685 | 0.4935 | |
| Regional-history product under the same 320-slot budget | |||
| Logical regions | 160 | 40 | one quarter as many |
| Mixed-group fraction | ||||
|---|---|---|---|---|
| Width | Depth 1 | Depth 2 | Depth 3 | Logical regions |
| 1 | 0.0000 | 0.0000 | 0.0000 | 320 |
| 2 | 0.0000 | 0.0000 | 0.0000 | 160 |
| 3 | 0.1961 | 0.2589 | 0.2808 | 106 |
| 4 | 0.1871 | 0.2459 | 0.2661 | 80 |
| 8 | 0.4547 | 0.5488 | 0.5805 | 40 |
| Depth | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
|---|---|---|---|---|---|---|---|
| Map acc., measured two-bit field | 97.2 | 95.0 | 93.1 | 91.3 | 90.0 | 88.1 | 86.9 |
| Map acc., 3-bit upper bound (106 regions) | 93.4 | 91.5 | 90.6 | 90.3 | 90.3 | 89.9 | 89.9 |
| Map acc., 4-bit upper bound (80 regions) | 90.6 | 88.1 | 86.9 | 85.6 | 86.3 | 85.6 | 85.3 |
| Map acc., 8-bit upper bound (40 regions) | 84.4 | 80.6 | 78.8 | 78.1 | 78.1 | 77.5 | 76.9 |
| mIoU, measured two-bit field | 93.9 | 87.3 | 79.0 | 72.5 | 67.1 | 62.8 | 59.0 |
| mIoU, 3-bit upper bound (106 regions) | 86.4 | 79.8 | 76.0 | 73.8 | 72.1 | 70.4 | 69.4 |
| Watermark | Reissuance | Evidence for an unissued output | Regional-history readout |
|---|---|---|---|
| VINE-R | Rewrite the generation-conditioned issue message over the complete image. | Message survival yields an issue match; message loss yields Open . | Head label at every cell. |
| TrustMark-Q | Rewrite the same issue message with the released quality-oriented TrustMark model. | Message survival yields an issue match; message loss yields Open . | Head label at every cell. |
| TrustMark-RM | Apply the released remover before rewriting the next generation-conditioned TrustMark message. | The rewritten message has the same two outcomes. | Head label at every cell. |
| OmniGuard | Rewrite the issue message with the released robust message watermark. | The released tamper-mask output supports Open ; message loss yields Open . | Head label at every cell; the snapshot mask carries no earlier generations. |
| WAM | A contiguous generation block and a 160-cell field in GSF’s layout; reissue rewrites the block and the committed cells. | An unreadable generation block or a disrupted cluster supports Open ; a field lost entirely or disrupted beyond one cluster abstains. | Generation code decoded on cells with readable WAM evidence. |
| GSF | Lineage evidence, generation-conditioned issue evidence, live head, and inherited keyed field. | Once the image is attributed to the lineage, undecodable issue evidence or a disrupted cluster supports Open without asserting an issue. | Spatial reader recovers the last authorized generation at every field cell. |
| Current / Superseded accuracy (%) | |||||
|---|---|---|---|---|---|
| System | Overall | ||||
| VINE-R | 99.8/– | 98.3/99.8 | 97.1/99.1 | 97.7/98.4 | 98.2/98.9 |
| TrustMark-Q | 100.0 /– | 100.0/100.0 | 100.0/100.0 | 99.8/100.0 | 99.9/100.0 |
| TrustMark-RM | 100.0 /– | 99.6/100.0 | 99.6/99.8 | 99.4/99.7 | 99.6/99.8 |
| OmniGuard tamper mask | 93.8/– | 95.2/93.8 | 96.5/94.5 | 95.8/95.1 | 95.3/94.7 |
| WAM in GSF layout † | 81.5/– | 0.2/81.5 | 0.0/40.8 | 0.0/27.2 | 20.4/40.8 |
| System | Open | False issue | Unknown |
|---|---|---|---|
| VINE-R | 63.5 [61.7, 65.3] | 36.5 [34.7, 38.3] | – |
| TrustMark-Q | 70.1 [68.4, 71.7] | 29.9 [28.3, 31.6] | – |
| TrustMark-RM | 70.5 [68.8, 72.2] | 29.5 [27.8, 31.2] | – |
| OmniGuard tamper mask | 61.3 [59.5, 63.1] | 38.7 [36.9, 40.5] | – |
| WAM in GSF layout † | 50.8 [48.9, 52.6] | 0.4 [0.2, 0.7] | 48.8 [47.0, 50.7] |
| Full GSF (ours) | 94.0 [93.0, 94.8] | 2.1 [1.6, 2.7] | 3.9 [3.2, 4.7] |
| Native | |||
|---|---|---|---|
| System | own / later / Open / false issue | own / later / Open / false issue | own / later / Open / false issue |
| VINE-R | 98.2 / 98.9 / 63.5 / 36.5 | 75.7 / 98.9 / 68.9 / 7.8 | 94.2 / 98.9 / 68.9 / 10.8 |
| TrustMark-Q | 99.9 / 100.0 / 70.1 / 29.9 | 79.8 / 100.0 / 72.0 / 7.1 | 97.6 / 100.0 / 72.0 / 10.6 |
| TrustMark-RM | 99.6 / 99.8 / 70.5 / 29.5 | 79.9 / 99.8 / 72.4 / 7.1 | 97.6 / 99.8 / 72.4 / 10.2 |
| OmniGuard tamper mask | 95.3 / 94.7 / 61.3 / 38.7 | 67.4 / 94.7 / 65.8 / 7.7 | 90.1 / 94.7 / 65.8 / 11.3 |
| Registry key | Clean: Current / Superseded | Clean: false issue | Delivered: Current retention |
|---|---|---|---|
| Exact hash | 100.0 / 100.0 | 0.0 | 0.0 |
| Perceptual hash, | 100.0 / 99.7 | 0.5 | 83.1 |
| Perceptual hash, | 100.0 / 99.7 | 14.8 | 100.0 |
| Perceptual hash, | 100.0 / 99.7 | 31.1 | 100.0 |
| Perceptual hash (oracle) | 100.0 / 99.7 | 2.6 | 99.4 |
| GSF (reads no per-issue record) | 99.5 / 99.8 | 2.2 | 98.2 |
| Where the map comes from | Map acc. | mIoU |
|---|---|---|
| Lookup alone, no edit record read: every cell gets the identified issue | 56.3 | 14.3 |
| Lookup replay of every issue’s committed scope and parent link | 100.0 | 100.0 |
| GSF: read from the copy, with the key and the list of issues | 93.1 | 79.0 |
| Read from | Map acc. | mIoU |
|---|---|---|
| , before the edit | 93.6 | 79.7 |
| Paste-back edit of | 93.7 | 79.4 |
| Carrier (initial-issue PSNR) | |||
|---|---|---|---|
| WAM in GSF layout, committed cells only (Table 3 ; 37.8 dB) | 49.05 | 32.38 | 24.30 |
| WAM in GSF layout, full refresh (46.7 dB) | 63.61 | 43.56 | 34.22 |
| WAM in GSF layout, full refresh (37.8 dB) | 80.40 | 55.68 | 43.15 |
| GSF (46.7 dB) | 93.87 | 87.27 | 78.96 |
| Full refresh, 46.7 dB | Full refresh, 37.8 dB | GSF | |||
| Frozen rules | Tuned | Frozen rules | Tuned | ||
| Readout tolerance (bits) | 4 | 10 | 4 | 8 | – |
| 1 | 64 | 1 | 64 | 1 | |
| 1 | 256 | 1 | 1 | 1 | |
| 1/4 | 1/7 | 1/4 | 1 | 1/4 | |
| Current | 6.9 | 90.7 | 21.0 | 90.9 | 99.5 |
| Query | Issue evidence | Identifier | Decision | |
|---|---|---|---|---|
| Unmarked image, other content | 480 | 0 | 0 | Unknown 480 |
| Unmarked source of the named lineage | 480 | 0 | 0 | Unknown 480 |
| Issued of another lineage | 4,320 | 0 | 0 | Unknown 4,320 |
| of the named lineage (control) | 480 | 480 | 480 | Current 475, Open 2, Unk. 3 |
| Experiment | GSF | Reference | |
|---|---|---|---|
| Object regenerated and pasted back | 960 | 0.0 | TrustMark-Q: 38.8 |
| Unissued output pasted into the issue | 1,920 | 3.3–5.8 | TrustMark-Q: 68.5–98.5 |
| Further editors | 2,400 | 0.0 | – |
| Removal | 2,400 | 0.0 | – |
| Transplanted issue evidence | 1,920 | 0.0–3.3 | – |
| Copied watermark | 1,920 | 0.4–8.3 | global-message watermarks: 96.5–99.8 |
| Attack | read Current | Unissued edits accepted as issues |
|---|---|---|
| No attack | 475 of 480 | 34 of 1,920 (0.0–2.5% by kind) |
| Removal (five operators) | 1 of 2,400 | 0 of 2,400 editor outputs |
| 3 3 median filter | 96.3% | – |
| Released watermark over | 90.6–98.8% | – |
| Second GSF watermark under another key | 0 of 480 | – |
| Transplanted issue evidence | – | 32 of 1,920 (0.0–3.3%) |
Data Provenance
Data provenance tracks where data and AI outputs come from, how they are transformed, and what evidence supports their attribution and use.
See all 205 papers