WorksComputer Science
TetherMark: Instance-Level Watermark Authentication Against Cross-Instance Transplantation
Abstract
Robust watermarking is commonly verified by asking whether a hidden payload remains decodable after compression, geometric distortion, or generative editing. Localized and semantic extensions may additionally check whether an expected region or category is still present. These tests establish evidence validity or content consistency, but not whether the evidence still belongs to the particular visual instance registered by its owner. We expose cross-instance transplantation: an attacker extracts or estimates valid watermark evidence from one registered image and transfers it to a different image of the same category. The message can still decode, the expected content can remain present, and the result can look natural, even though the evidence has been reassigned to the wrong instance. To enable instance-level watermark authentication, we introduce TetherMark, a post-hoc watermark that combines two complementary checks. The message check verifies an edit-robust payload, while the relation check uses secret patch-pair selection, a registered response, natural correspondence transport, and enrollment normalization to test consistency with the registered instance. The verifier accepts only when both checks pass. Across 240 test subjects, TetherMark attains 92.08% legitimate true acceptance with 9.17% and 16.25% attack success under WmForger and WMCopier, respectively, yielding the highest instance-binding balanced accuracy under both attacks.
Introduction
Most watermark-based provenance systems frame verification as a robustness test: can the registered payload still be recovered after the image has been transformed? Traditional detectors therefore make a message-level decision. If the registered bits survive compression, geometric distortion, or generative editing, the query is accepted as watermarked [4, 12, 17, 25]. This confirms that watermark evidence is recoverable, but it does not compare the query with the visual instance for which that evidence was issued.
Content-aware watermarking brings verification closer to the protected content, but usually at a coarser level. Localized schemes check whether a designated region remains, while semantic approaches check whether the query still contains the expected category, caption, or concept [2, 15, 23, 37]. A query can therefore pass once it still contains, for example, a watch in the protected region; the check need not establish that it is the same watch. Payload validity and category/region consistency are thus distinct from the finer-grained question of whether valid evidence still belongs to the enrolled instance.
Consider a marketplace that registers a photograph of one specific wristwatch. An attacker estimates the watermark residual from that photograph and places it on a different wristwatch of the same product category. The traditional detector recovers the exact registered message and accepts. A semantic or localized auxiliary verifier also sees a watch in the expected subject region and accepts. The forged image can remain visually natural, yet the registered evidence now accompanies the wrong physical item. We call this failure cross-instance transplantation and study its hardest same-category case. Our objective is instance-level watermark authentication: accept legitimate edits of the registered watch while rejecting a different watch with copied evidence.
To this end, we introduce TetherMark, an instance-level authentication framework that explicitly separates evidence validity from registered-instance consistency. The message check verifies that at least one robust payload branch carries the registered message. The relation check verifies a private enrollment relation: the key selects patch pairs inside the subject, enrollment records their response, natural correspondences transport the same pairs through an edit, and enrollment normalization tests whether the aligned response remains consistent with the registered instance. The final decision is an AND rule. Thus a copied message may pass the first check, while the enrollment-relation check rejects its reassignment to a different instance.
Unlike Rel-Zero [6], which uses intrinsic relations as a stand-alone zero-watermark, TetherMark makes a keyed, enrollment-relative relation the registered-instance check of a post-hoc watermark authenticator. Across 240 test subjects and 2,400 legitimate edits, TetherMark attains a 92.08% true acceptance rate (TAR) and limits same-category attack success rate (ASR) to 9.17%/16.25% under WmForger/WMCopier, yielding the highest instance-binding balanced accuracies (91.46%/87.92%). Under 1,000-step key-unknown transfer, ASR remains 8.33%; on unseen categories, TAR is 84–92% with 4% copy ASR. Together, these results make message validity and relation consistency necessary for instance-level acceptance.
Our contributions are threefold:
We formulate cross-instance transplantation as a distinct threat and evaluation target: valid payload and category evidence must not authenticate the wrong instance.
We propose an instance-level watermark authenticator that accepts only when a robust message and a keyed, enrollment-relative subject relation are both valid; natural correspondence preserves it across legitimate edits.
We define a calibrated verification protocol and evaluate the mechanism with public-partition, no-shaping, enrollment-normalization, key-unknown transfer, unseen-category, and cost controls.
Related Work
Robust payload watermarking. Classical methods embed spread-spectrum or perceptually masked transform-domain signals [3, 8], while learned encoder–decoders train payload recovery under distortions, physical recapture, unknown corruptions, adversarial training, and JPEG compression [13, 18, 27, 29, 40]. Diffusion-era schemes root evidence in model parameters or generative latents [7, 11, 30, 31, 33, 35]. Post-hoc systems such as TrustMark, Robust-Wide, VINE, and Pixel Seal further improve recovery under arbitrary-resolution, instruction-driven, diffusion-based, or high-resolution edits [4, 12, 17, 25]. Their verifier remains message-centric: a decoded payload establishes evidence presence, while its assignment to the registered image instance remains unchecked.
Localized and semantic binding. Content-aware methods more closely associate evidence with what remains in the query. WAM localizes evidence and WOFA distributes it against partial theft [15, 23]; SWIFT, SEAL, and SemBind bind evidence to captions or semantics [2, 10, 37]. MetaSeal uses content-dependent cryptographic watermarks [39], while Zhu et al. target content watermarking against removal and forgery [41]. These designs verify stronger content consistency than a global payload alone. A different instance can still share the same category, protected region, caption, or prompt. Instance-level authentication therefore requires a registered-instance consistency test.
Watermark forgery and transplantation. Recent attacks separate detector acceptance from legitimate provenance. Black-box attacks transfer semantic evidence across diffusion generators [20]; WmForger optimizes from one marked example [26]; WMCopier injects a learned marked-image signal through shallow inversion [9]; WForge recovers transferable watermark-like residuals [28]. WAVES standardizes distortion, regeneration, and adversarial tests [1], while regeneration and averaging expose removal and reusable-pattern failure modes [32, 38]. These studies establish that watermark evidence can be forged, transferred, or removed. Their success criteria measure evidence manipulation broadly; our evaluation isolates the same-category reassignment of a valid payload while expected content remains present.
Relational authentication. Rel-Zero [6] registers naturally stable patch-pair relations as a zero-watermark for generative editing. TetherMark extends relational evidence with a conventional robust payload and a separately evaluated registered-instance check. Secret pair selection and enrollment normalization form the relation check, correspondence transports it to the query, and the AND rule combines it with message validity. Active shaping further separates the registered response.
Problem Formulation and Threat Model
Enrollment and legitimate transformations. The owner holds an image , a subject mask , a payload , and a secret key . Enrollment produces a marked image and a private record describing the protected instance. Given a query image , the verifier receives and must determine whether is a legitimate transformation of the enrolled instance. We allow changes to the background, illumination, material, style, viewpoint, and moderate generative appearance, provided that provenance continuity of the enrolled subject is preserved. Thus, a stylized rendering of an enrolled watch may be accepted, whereas replacing it with an unrelated watch should not be, even if both images remain semantically consistent. The private record contains the registered source-subject cache and all information needed for verification; the edit prompt and diffusion seed remain outside the interface.
The unknown transfer from white-box optimization against a public subject verifier.
Cross-instance transplantation. In contrast to a legitimate transformation, a transplantation attack extracts or estimates valid watermark evidence from registered instance and places it on a different instance . The adversary knows the algorithms and public feature extractors, obtains at least one marked image, and may copy or estimate its watermark residual. The adversary can apply generative editing and optimize a bounded perturbation against a public identity matcher. Our primary case chooses from the same semantic category as : message validity, category consistency, and visual naturalness can all be preserved while the evidence–instance assignment is wrong. The key and enrollment record remain secret. We additionally test key-unknown transfer from white-box optimization against a public subject verifier.
Instance-level authentication objective. The message check asks whether registered watermark evidence is present; the relation check asks whether that evidence remains consistent with the enrolled instance. The final decision is
Our primary evaluation targets same-category cross-instance transplantation, where payload and semantic checks remain plausible. Unmarked images, wrong keys, and subject replacement form auxiliary rejection cases around this instance-assignment test.
Method
TetherMark implements Eq. (1) through the three stages in Figure 1. Embedding and enrollment fuse the VINE/WAM message carrier with a keyed group-relation carrier and save the private record. Relation readout aligns the registered source and group masks with a query, then converts their RGB residual into relation logits. Instance-level authentication evaluates the decoder OR path and relation-gate path in parallel, followed by the final AND decision.
Figure 1. Overview of the TetherMark workflow. Embedding and enrollment. VINE and WAM residuals form the message carrier. Mask-filtered Rel-Zero tokens define four registered groups; a fixed carrier, the keyed balanced codeword, and the relation writer produce the relation residual. Relation readout. After source–query alignment, the RGB residual and registered group masks pass through the CNN group reader and cyclic difference to obtain relation logits. Instance-level authentication. An OR rule combines the two payload decoders, while codeword scores and the enrollment reference feed the relation gate. An AND rule accepts only when both paths are valid.
Embedding and Enrollment
The Embedding and Enrollment stage in Figure 1 takes image , subject mask , payload , and secret key . VINE and WAM encode their registered branch payloads and their residuals are summed into the message carrier. Inside , frozen Rel-Zero tokens are filtered by mask coverage and clustered into four registered group masks. The key selects ordered group/patch pairs and a balanced codeword; starting from a fixed carrier, the relation writer shapes the corresponding low-amplitude state. Enrollment then records the response observed in the released image, the group geometry, and correspondence data. Thus this stage releases and registers .
Let and be the VINE and WAM residuals for , respectively. For a candidate relation perturbation , define
Here is obtained below. Both branches contribute to the same protected image, with the relation perturbation restricted to . The message branch provides payload capacity and edit robustness. The relation branch supplies the keyed response used for instance-level consistency, while enrollment anchors that response to the registered source.
Let denote the registered group masks. The secret key selects relation components and ordered group/patch pairs,
where index supports in and selects a component of the shared CNN group reader. Let denote the source-aligned RGB residual at enrollment, and let be the reader descriptor of residual restricted to support . The response associated with each tuple is
The ordered difference makes the response depend on a relation between two patches. The key also samples a balanced code with . Complementary signs discourage a common-direction response shortcut.
With four components, six balanced patterns are available; our evaluated one-bit instantiation uses one complementary pair and treats the other patterns as wrong-key alternatives. Pair and component choices provide further key diversity.
The fixed carrier shown in Figure 1 initializes ; the encoders, Rel-Zero extractor, and relation reader remain frozen while the writer optimizes its pixels. We write the code on the combined image :
The quality loss combines pixel and perceptual distortion penalties against the original image. Including in makes relation shaping account for the image actually released. The budget limits the additional perturbation; payload recovery must also be checked on the resulting because both branches share the carrier. § 5.3 separately tests whether shaping is necessary once the private pair selection, registered response, correspondence transport, and enrollment-normalized gate are retained.
After substituting into Eq. (2), we evaluate Eq. (4) on and register
Here is the registered source-subject cache and is the reference response read from the released image. stores frozen tokens, while stores the four group masks and their subject-relative geometry for alignment. The correspondence descriptors can differ from the relation reader ; no paired opposite-code counterpart is required.
Relation Readout
During the Relation Readout stage in Figure 1, the verifier receives query and enrollment record . Because generative editing can move, reshape, or partially redraw a subject, we use from Eq. (6) to locate registered patches in the query. Frozen grayscale token descriptors provide mutual nearest-neighbor matches between the registered source and . Let map a matched source location to its query location, and let denote the source cache transported into the query frame. We transport the registered masks in by the same map. The subtraction node in Figure 1 forms the aligned RGB residual . The shared CNN group reader processes on each transported mask. Its cyclic-difference block implements the ordered subtraction for every registered pair:
Thus, the displayed relation logits measure the same ordered group relations as the enrolled response . Frozen descriptor similarity establishes correspondence; the keyed CNN readout supplies the authentication evidence.
We denote correspondence validity by . It requires reliable matches for the selected pair endpoints and sufficient subject match coverage and confidence, with thresholds fixed using only the 60 calibration subjects. If , the query is rejected before constructing or evaluating the relation gate. This rejection rule is included explicitly in the verifier below.
Instance-Level Authentication
The Instance-Level Authentication stage in Figure 1 evaluates two paths in parallel. VINE and WAM decode the query independently; the OR node sets MessageValid when either branch passes its own frozen threshold. We denote the aggregated result by . On the relation path, directly gives RelationValid ; for , the codeword-score block compares in Eq. (7) with the registered response . Raw relation magnitude varies across subjects and edit strengths, so a large response alone does not establish agreement with enrollment. We therefore feed both raw statistics and enrollment-relative features to a linear gate:
Semicolons denote concatenation; is the sorted response vector, and are its mean and standard deviation. The ratio is elementwise, with for stability. We use and the projection coefficient .
The “Codeword Scores” block in Figure 1 produces
Here is a sampled set of wrong keys. We regenerate their codes and pairs as in § 4.1, then recompute their responses with Eq. (7) on retained enrollment tokens. Consequently, compares independently sampled wrong-key responses. The sampling protocol is fixed before test evaluation.
The displayed “Logistic Gate” is a calibrated linear map from Eq. (8) to the relation score. Including correspondence rejection gives
where is standardized using means and scales estimated exclusively from the 60 calibration subjects, and includes a bias coefficient. The gate parameters and its decision threshold are fitted only on allowed edits and cross-instance transplants from these calibration subjects. All correspondence thresholds, feature statistics, gate parameters, key-sampling protocol, and authentication thresholds are then frozen. The remaining 60 development subjects are used only to validate this frozen rule and do not contribute to feature standardization, gate fitting, or threshold selection. The final test subjects are used only for evaluation.
Finally, let denote the possibly aggregated payload-branch confidence; the evaluated aggregation is specified below. For frozen thresholds and , Eq. (1) becomes
The first indicator is the decoder OR path and tests whether registered evidence is present. The second is the relation-gate path, including correspondence validity, and tests consistency with the enrolled instance. Their product is the final AND node in Figure 1. A transplanted payload can pass the first path while failing the second. The key-unknown transfer experiment measures this separation against optimization through a public verifier.
Implementation Details
Partition and writing. We retain RGB within the subject mask, fill its exterior with 127/255, and resize the subject bounding-box crop to . The frozen Rel-Zero student ViT [6] yields a grid of -normalized tokens. We require at least eight patches with subject occupancy of at least 0.25. Spherical -means uses farthest-point initialization and at most 20 iterations to produce the four groups. We cap the merged carrier at an ROI RMS of and an norm of , then upsample it from to and apply three rounds of RMS restoration and clipping. The fusion weight gives nominal relation-component bounds of (ROI RMS) and ().
Readout and decision. The shared CNN group reader applies four stride-two convolutions to each registered-mask residual, then forms the cyclic pair differences in Eq. (7). The convolutions use kernel sizes and output widths . Each convolution is followed by GELU. Adaptive average pooling and a linear layer produce a scalar. We fit the relation gate with sample-weighted logistic regression using an penalty of 1.0 on the coefficients and an unpenalized intercept. Newton/IRLS runs for up to 100 iterations and stops when the largest absolute parameter update falls below . In our redundant payload branch, MessageValid holds when either at least 71 of the 100 VINE-B bits match or all 32 WAM bits match.
Experiments
We organize the evaluation around one question: can the conjunction of message validity and a registered-instance relation turn evidence detection into instance-level watermark authentication? TAR tests whether both checks survive legitimate edits of instance ; ASR tests whether a same-category instance carrying copied evidence is incorrectly accepted under ’s enrollment. We compare two transplantation attacks, break down TAR by edit type, and ablate the components of the relation check. We then test key-unknown transfer from a public subject verifier, generalization to unseen categories, and the fidelity, storage, and runtime cost of instance binding.
Experimental Setup
Data and protocol. We use Stanford Online Products [24], with 120 development and 240 test subjects from disjoint product groups across six categories. We pre-partition the development cohort into 60 calibration subjects and 60 held-out subjects. We fit the relation gate and select its threshold on the calibration split, then validate the frozen rule on the held-out split. We obtain subject masks with Grounding DINO [16] and GrabCut [22]. The ten edit conditions cover compression, resizing, compositing, and generative edits from Stable Diffusion 1.5 Inpainting [21]. A generative-edit pair is eligible when Grounding DINO detects the enrolled category with a score of at least 0.10 in both the unmarked and marked subject crops. All ten conditions yield 240 eligible queries each. WmForger and WMCopier share 240
fixed same-category source–target pairs drawn from distinct products. Every transplant is scored against the source enrollment, so attack success requires the verifier to assign copied evidence to the wrong instance.
Baselines. We compare with TrustMark-P [4], Pixel Seal [25], Robust-Wide [12], WAM [23], SWIFT [10], and VINE [17]. We also evaluate VINE with a learned DINO/SigLIP subject verifier that uses an oracle query mask. This baseline requires both its message and subject checks to pass. TetherMark requires its relation check and at least one of the VINE and WAM message checks. Each method uses its own operating point, fixed before testing.
Metrics. True acceptance rate (TAR) is the fraction of eligible edits accepted by the full verifier; condition-specific TAR uses the eligible queries for that edit, while overall TAR pools the 2,400 queries across all ten conditions. For each attack setting, attack success rate (ASR) is the fraction of all scheduled attempts in which a different instance carrying copied evidence is incorrectly accepted under the source enrollment; outputs that fail the shared validity audit count as unsuccessful attacks. For attack , instance-binding balanced accuracy (IBA) measures the two sides of instance-level watermark authentication together—accepting legitimate edits and rejecting the wrong instance: . All three rates are reported as percentages. In the key-unknown transfer test, coverage is the percentage of generated outputs that pass the shared validity audit, regardless of verifier acceptance.
For fidelity, peak signal-to-noise ratio (PSNR, dB; higher is better) measures pixel-level agreement between each clean and marked image, whereas LPIPS-Alex [36] (lower is better) measures their learned perceptual distance using the AlexNet backbone [14]. Table 6 reports medians over marked images. PNG size is the encoded marked-image size, enrollment size is the on-disk per-subject record excluding shared model weights, and embed/detect latency is CUDA-synchronized milliseconds per image under the hardware and timing window specified in § 5.6.
| Method | Payload configuration | Median PSNR (dB) ↑ | Median LPIPS-Alex ↓ | Median PNG size | Per-subject enrollment record | Embed / detect (ms) |
| TrustMark-P | 100 bit | 43.5 | 0.0027 | 255,000 B | No source image | 45.1 / 25.0 |
| Pixel Seal | 256 bit | 48.9 | 0.0013 | 248,000 B | No source image | 9.13 / 12.0 |
| Robust-Wide | 64 bit | 41.91 | 0.0085 | 260,000 B | No source image | 25.0 / 18.0 |
| WAM | 32-bit message | 40.09 | 0.0202 | 258,527 B | 0 B | 9.64 / 10.41 |
| SWIFT | Semantic vector (60-bit equiv.) | 41.0 | 0.0090 | 265,000 B | Semantic commitment No source image | 30.0 / 22.0 |
| VINE | 100-bit message | 34.19 | 0.0286 | 251,936 B | 0 B | 60.53 / 5.58 |
| VINE + Learned Subject Check | 100-bit message + subject check | 34.19 | 0.0286 | 251,936 B | 65,000 B (features) | 60.53 / 15.0 |
| TetherMark (ours) | 100-bit VINE + 32-bit WAM + 1-bit relation state | 32.90 | 0.0533 | 292,737 B | 614,384 B1 | 76.03 / 18.82 |
Table 6. Payload configuration, image fidelity, storage, and runtime. VINE and WAM use globally registered expected messages.
Transplantation Resistance and Edit Robustness
TetherMark achieves 92.08% TAR and ASRs of 9.17% under WmForger and 16.25% under WMCopier (Table 1), giving the highest IBA for both attacks. Compared with VINE with learned subject checking, its TAR is 28.45 percentage points higher and its ASRs are 76.25 and 78.75 points lower, respectively. TetherMark simultaneously accepts more legitimate edits and rejects more wrong-instance transfers. The result supports the joint message–relation rule: copied payloads satisfy the message branch, while the source enrollment rela-tion determines instance-level authenticity. Compared with
| Method | Legit. TAR ↑ | WmForger ASR ↓ / IBA ↑ | WMCopier ASR ↓ / IBA ↑ |
| TrustMark-P | 89.38 | 80.00 / 54.69 | 90.83 / 49.27 |
| Pixel Seal | 89.75 | 84.58 / 52.58 | 85.42 / 52.17 |
| Robust-Wide | 80.00 | 100.00 / 40.00 | 99.17 / 40.42 |
| WAM | 77.46 | 83.33 / 47.06 | 95.83 / 40.81 |
| SWIFT | 74.38 | 77.92 / 48.23 | 89.58 / 42.40 |
| VINE | 69.96 | 84.58 / 42.69 | 97.50 / 36.23 |
| VINE + Learned Subject Check | 63.63 | 85.42 / 39.10 | 95.00 / 34.31 |
| TetherMark (ours) | 92.08 | 9.17 / 91.46 | 16.25 / 87.92 |
Table 1. Same-category cross-instance transplantation. TAR pools 2,400 edited queries. Each attack column reports ASR / IBA over 240 attempts. All values are percentages.
VINE, the learned subject-checking baseline has a TAR that is 6.33 percentage points lower, while its WMCopier ASR is 2.50 points lower. We generate the WmForger attack images for the two VINE variants separately, using the same source–target pairs.
Table 2 reports TAR by edit condition. TetherMark retains 100% TAR after JPEG-70 compression and 0.5× resizing, and 88.75–93.75% after cut-and-paste and background replacement. On these three compositing conditions, both VINE variants retain 2.08% TAR and Pixel Seal retains 92.08–99.58%. On the five generative edits, VINE and TrustMark-P attain 97.50–99.58% and 94.17–95.83% TAR, respectively, compared with 85.00–92.92% for TetherMark. The compositing and generative-edit results show that the relation check follows the enrolled subject across spatial and appearance changes. The remaining gap quantifies the robustness–binding trade-off.
| Method | JPEG-70 | 0.5× scale | Cut-and-paste | Cool bg. | Warm bg. | Material | Watercolor | Technical style | Redraw 0.25 | Redraw 0.35 |
| TrustMark-P | 95.83 | 100.00 | 75.00 | 72.92 | 74.17 | 95.83 | 95.00 | 94.17 | 95.83 | 95.00 |
| Pixel Seal | 100.00 | 100.00 | 99.58 | 95.00 | 92.08 | 80.83 | 83.33 | 80.42 | 84.17 | 82.08 |
| Robust-Wide | 97.92 | 99.58 | 80.42 | 78.33 | 79.58 | 72.08 | 71.25 | 70.00 | 75.42 | 75.42 |
| WAM | 81.67 | 100.00 | 94.58 | 95.42 | 92.92 | 62.92 | 61.67 | 62.92 | 60.83 | 61.67 |
| SWIFT | 90.00 | 95.00 | 75.42 | 73.33 | 74.58 | 67.08 | 65.83 | 65.00 | 69.17 | 68.33 |
| VINE | 100.00 | 100.00 | 2.08 | 2.08 | 2.08 | 98.33 | 98.75 | 99.17 | 99.58 | 97.50 |
| VINE + Learned Subject Check | 99.58 | 99.58 | 2.08 | 2.08 | 2.08 | 84.58 | 82.50 | 85.83 | 95.42 | 82.50 |
| TetherMark (ours) | 100.00 | 100.00 | 93.75 | 93.75 | 88.75 | 85.42 | 90.00 | 91.25 | 92.92 | 85.00 |
Table 2. Robustness to legitimate edits. TAR (%) under image processing, compositing, and generative editing. Each condition contains 240 eligible queries. Redrawing uses denoising strengths of 0.25 and 0.35.
Ablation of the Relation Verifier
For each variant, we evaluate one material edit per test subject, yielding 240 legitimate queries and 240 attempts per attack. The variants share the subjects, masks, messages, keys, editor settings, prompts, seeds, and source–target pairs. We regenerate edits and attacks for the two carrier variants, which disable active relation shaping or replace the private partition with a public grid. Both use the full model’s frozen decision rule. The variant without enrollment normalization reuses the full model’s images and refits its relation readout on the same 60 calibration subjects used for the full model (§ 5.1). For both readouts, we freeze the highest threshold that retains at least 90% TAR in every valid calibration edit condition before testing.
The full model achieves the highest IBA under both attacks (Table 3). Removing active relation shaping lowers material TAR by 2.92 percentage points. It increases WmForger ASR by 0.83 points and decreases WMCopier ASR by 2.50 points. Replacing the private partition with a public grid yields the lowest ASR under both attacks, but reduces material TAR by 27.92 points. Removing enrollment normalization reduces material TAR by 5.42 points and increases ASR by 14.58 points under WmForger and 12.08 points under WMCopier.
| Variant | Material TAR ↑ | WmForger ASR ↓ / IBA ↑ | WMCopier ASR ↓ / IBA ↑ |
| w/o active relation shaping | 82.50 | 10.00 / 86.25 | 13.75 / 84.38 |
| w/ public partition | 57.50 | 2.08 / 77.71 | 3.75 / 76.88 |
| w/o enrollment normalization | 80.00 | 23.75 / 78.13 | 28.33 / 75.83 |
| TetherMark (full) | 85.42 | 9.17 / 88.13 | 16.25 / 84.58 |
Table 3. Ablation of the relation verifier. Material TAR uses one eligible edit per subject. ASR uses 240 attempts per attack. IBA is the mean of material TAR and the corresponding attack rejection rate. All values are percentages.
Key-Unknown Transfer Attack
The attacker has white-box access to the public subject verifier in VINE with learned subject checking. Both verifiers score the resulting images at each budget, with TetherMark evaluated offline for transfer. The attacker knows the source enrollment image, the VINE message, and the public verifier’s parameters and threshold. The TetherMark relation key, private partition, and enrollment statistics remain private. We hash each test subject ID with the fixed prefix relbind-v2-whitebox60: using SHA-256, take the first 60 subjects in hash order, and retain their preassigned same-category targets. For each pair, both budgets start from the same exact residual transplant. The attack uses sign-gradient PGD [19] to maximize the public verifier’s frozen logit, which combines standardized Rel-Zero/DINO patch statistics [5] and SigLIP crop similarities [34]. PGD restricts updates to the oracle target mask, with an budget of 4/255 relative to the initial residual transplant, a step size of 0.5/255, and random seed 512. We generate outputs separately for 100 and 1,000 gradient steps per pair.
Table 4 reports ASR over all 60 attempts per budget. All outputs pass the shared validity audit for successful genera-
| PGD steps | VINE + subject ASR ↓ | TetherMark ASR ↓ | Coverage (%) |
| 100 | 40.00 | 8.33 | 100.00 |
| 1,000 | 56.67 | 8.33 | 100.00 |
Table 4. White-box PGD against VINE with learned subject checking and transfer to TetherMark. Coverage is the percentage of attack outputs that pass the shared validity audit.
tion, target presence, mask support, and perturbation budget. Increasing the optimization horizon from 100 to 1,000 steps raises the learned baseline’s ASR from 40.00% to 56.67%. TetherMark accepts 5 of the 60 attack images (8.33%) at each budget. The persistent gap at 1,000 steps shows weak transfer from the public subject verifier to the private relation key, partition, and enrollment record. The experiment directly measures transfer into the relation check.
Generalization to Unseen Categories
We test 25 held-out sofa and 25 toaster subjects. Redraw uses Stable Diffusion 1.5 Inpainting (strength 0.35, 30 steps, guidance 7.5) with matched seeds and an identity-preserving prompt. Within each category, the next subject by ID is the copy target, giving 50 fixed pairs. We add each TetherMark source residual to its clean target and let all three verifiers score the same copies. Table 5 shows the trade-off: TetherMark keeps 84–92% TAR, 10–20 points above the learned subject baseline, while accepting only 2/50 copies (4% ASR), versus 11/50 for VINE. The learned baseline rejects all copies but accepts four fewer legitimate edits.
| Condition | VINE | VINE + subject | TetherMark |
| Material TAR | 98% | 72% | 92% |
| Watercolor TAR | 100% | 68% | 86% |
| Technical-style TAR | 98% | 68% | 84% |
| Redraw TAR | 96% | 76% | 86% |
| Copy ASR ↓ | 22% | 0% | 4% |
Table 5. Generalization to 50 subjects from unseen categories. The four edit columns report legitimate TAR. Copy ASR is measured on the same 50 residual copies for all three methods.
Image Fidelity, Payload, and Cost
Table 6 quantifies the cost of instance binding. Relative to VINE, TetherMark has lower median PSNR (32.90 vs. 34.19 dB), higher embed/detect latency (76.03/18.82 vs. 60.53/5.58 ms), and a 614,384 B median enrollment record.
VINE and WAM redundantly implement the message check; the evaluated relation branch selects one balanced four-component bit and audits the five alternative codewords. Timings use an RTX A6000, 512 × 512 inputs, batch size one, preloaded models, and CUDA synchronization. Detection is timed end to end after loading the enrollment record. Its 18.82 ms includes VINE/WAM decoding, frozen-token extraction and mutual-NN matching, RGB-residual reading, codeword scoring, and the logistic gate. Grounding DINO/GrabCut is enrollment-only and excluded, along with disk I/O and PNG encoding. These measurements characterize its storage and compute footprint.
Decision States
The two paths expose four outcomes. Valid message and relation checks accept a legitimate transformation. A message-only pass identifies decodable evidence assigned to the wrong instance, the signature of cross-instance transplantation. A relation-only pass indicates source-consistent content with lost payload evidence, while failure of both checks indicates that neither registered signal remains. These states localize each decision to payload recovery, correspondence, or enrollment consistency.
Enrollment fixes the source cache, group masks, keyselected pairs, reference response, and thresholds. Every query uses this record, and the verifier reports both branch outcomes with the final AND decision.
Conclusion
We introduced cross-instance transplantation, where valid evidence moves to a same-category target while payload and category checks remain valid. TetherMark accepts a query when its robust message and secret enrollment-bound relation agree. Correspondence transports the relation through legitimate edits, and enrollment normalization compares it with the registered response. Across same-category, key-unknown, and unseen-category tests, this conjunction preserves legitimate edits and rejects cross-instance transfers.
References
- [1]Bang An, Mucong Ding, Tahseen Rabbani, Aakriti Agrawal, Yuancheng Xu, Chenghao Deng, Sicheng Zhu, Abdirisak Mohamed, Yuxin Wen, Tom Goldstein, and Furong Huang. WAVES: Benchmarking the robustness of image watermarks. In International Conference on Machine Learning, 2024.arxiv.org/abs/2401.08573
- [2]Kasra Arabi, R. Teal Witter, Chinmay Hegde, and Niv Cohen. SEAL: Semantic aware image watermarking. In ICCV, pages 16196–16205, 2025.arxiv.org/abs/2503.12172
- [3]Mauro Barni, Franco Bartolini, and Alessandro Piva. Improved wavelet-based watermarking through pixel-wise masking. IEEE TIP, 10(5):783–791, 2001.DOI
- [4]Tu Bui, Shrujit Agrawal, and John Collomosse. TrustMark: Robust watermarking and watermark removal for arbitrary resolution images. In ICCV, 2025.
- [5]Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650–9660, 2021.arxiv.org/abs/2104.14294
- [6]Pengzhen Chen, Yanwei Liu, Xiaoyan Guo, Xiaojun Chen, Wu Liu, and Weiping Wang. Rel-zero: Harnessing patch-pair invariance for robust zero-watermarking against AI editing. In CVPR, pages 3337–3346, 2026.
- [7]Hai Ci, Pei Yang, Yiren Song, and Mike Zheng Shou. RingID: Rethinking tree-ring watermarking for enhanced multi-key identification. In ECCV, pages 338–354, 2024.
- [8]Ingemar J. Cox, Joe Kilian, F. Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE TIP, 6(12):1673–1687, 1997.pubmed.ncbi.nlm.nih.gov/18285237
- [9]Ziping Dong, Chao Shuai, Zhongjie Ba, Peng Cheng, Zhan Qin, Qinglong Wang, and Kui Ren. WMCopier: Forging invisible watermarks on arbitrary images. In NeurIPS, 2025.arxiv.org/abs/2503.22330
- [10]Gautier Evennou, Vivien Chappelier, Eva Kijak, and Teddy Furon. SWIFT: Semantic watermarking for image forgery thwarting. In IEEE International Workshop on Information Forensics and Security, 2024.arxiv.org/abs/2407.18995
- [11]Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In ICCV, pages 22466–22477, 2023.arxiv.org/abs/2303.15435
- [12]Runyi Hu, Jie Zhang, Ting Xu, Jiwei Li, and Tianwei Zhang. Robust-wide: Robust watermarking against instruction-driven image editing. In ECCV, 2024.arxiv.org/abs/2402.12688
- [13]Zhaoyang Jiao, Han Fang, and Weiming Zhang. MBRS: Enhancing robustness of DNN-based watermarking by mini-batch of real and simulated JPEG compression. In ACM MM, 2021.DOI
- [14]Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In NeurIPS, pages 1097–1105, 2012.
- [15]Gaozhi Liu, Silu Cao, Zhenxing Qian, Xinpeng Zhang, Sheng Li, and Wanli Peng. Watermarking one for all: A robust watermarking scheme against partial image theft. In CVPR, pages 8225–8234, 2025.DOI
- [16]Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024.arxiv.org/abs/2303.05499
- [17]Shilin Lu, Zihan Zhou, Jiayou Lu, Yuanzhi Zhu, and Adams Wai-Kin Kong. Robust watermarking using generative priors against image editing: From benchmarking to advances. In ICLR, 2025.arxiv.org/abs/2410.18775
- [18]Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR, pages 13548–13557, 2020.arxiv.org/abs/2001.04580
- [19]Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018.arxiv.org/abs/1706.06083
- [20]Andreas Müller, Denis Lukovnikov, Jonas Thietke, Asja Fischer, and Erwin Quiring. Black-box forgery attacks on semantic watermarks for diffusion models. In CVPR, 2025.arxiv.org/abs/2412.03283
- [21]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684–10695, 2022.arxiv.org/abs/2112.10752
- [22]Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. GrabCut: Interactive foreground extraction using iterated graph cuts. ACM TOG, 23(3):309–314, 2004.
- [23]Tom Sander, Pierre Fernandez, Alain Durmus, Teddy Furon, and Matthijs Douze. Watermark anything with localized messages. In ICLR, 2025.arxiv.org/abs/2411.07231
- [24]Hyun Oh, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4004–4012, 2016.
- [25]Tomáš Souček, Pierre Fernandez, Hady ElSahar, Sylvestre-Alvise Rebuffi, Valeriu Lacatusu, Tuan Tran, Tom Sander, and Alexandre Mourachko. Pixel seal: Adversarial-only training for invisible image and video watermarking, 2025.arxiv.org/abs/2512.16874
- [26]Tomáš Souček, Sylvestre-Alvise Rebuffi, Pierre Fernandez, Nikola Jovanović, Hady Elsahar, Valeriu Lacatusu, Tuan Tran, and Alexandre Mourachko. Transferable black-box one-shot forging of watermarks via image preference models. In NeurIPS, 2025.arxiv.org/abs/2510.20468
- [27]Matthew Tancik, Ben Mildenhall, and Ren Ng. StegaStamp: Invisible hyperlinks in physical photographs. In CVPR, pages 2117–2126, 2020.arxiv.org/abs/1904.05343
- [28]Yupu Wang, Zhengyuan Jiang, Yuepeng Hu, and Neil Zhenqiang Gong. Forging image watermarks by reversing watermark removal attacks. OpenReview, 2026. Submission to ICLR 2026.
- [29]Bingyang Wen and Sergul Aydore. ROMark: A robust watermarking system using adversarial training, 2019.arxiv.org/abs/1910.01221
- [30]Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Invisible fingerprints for diffusion images. In NeurIPS, 2023.DOI
- [31]Xun Xian, Ganghua Wang, Xuan Bi, Jayanth Srinivasa, Ashish Kundu, Minghui Hong, and Jie Ding. RAW: A robust and agile plug-and-play watermark framework for AI-generated images with provable guarantees. In NeurIPS, 2024.arxiv.org/abs/2403.18774
- [32]Pei Yang, Hai Ci, Yiren Song, and Mike Zheng Shou. Can simple averaging defeat modern watermarks? In NeurIPS, 2024.DOI
- [33]Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In CVPR, pages 12162–12171, 2024.arxiv.org/abs/2404.04956
- [34]Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975–11986, 2023.arxiv.org/abs/2303.15343
- [35]Lijun Zhang, Xiao Liu, Antoni Viros Martin, Cindy Xiong Bearfield, Yuriy Brun, and Hui Guan. Attack-resilient image watermarking using stable diffusion. In NeurIPS, 2024.arxiv.org/abs/2401.04247
- [36]Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586–595, 2018.arxiv.org/abs/1801.03924
- [37]Xin Zhang, Zijin Yang, Kejiang Chen, Linfeng Ma, Weiming Zhang, and Nenghai Yu. SemBind: Binding diffusion watermarks to semantics against black-box forgery attacks. In International Conference on Machine Learning, 2026.arxiv.org/abs/2601.20310
- [38]Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative AI. In NeurIPS, 2024.arxiv.org/abs/2306.01953
- [39]Tong Zhou, Ruyi Ding, Gaowen Liu, Charles Fleming, Ramana Rao Kompella, Yunsi Fei, Xiaolin Xu, and Shaolei Ren. MetaSeal: Defending against image attribution forgery through content-dependent cryptographic watermarks. Transactions on Machine Learning Research, 2026.arxiv.org/abs/2509.10766
- [40]Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. HiDDeN: Hiding data with deep networks. In ECCV, pages 657–672, 2018.arxiv.org/abs/1807.09937
- [41]Yifan Zhu, Yihan Wang, and Xiao-Shan Gao. Towards robust content watermarking against removal and forgery attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Findings, pages 8059–8069, 2026.arxiv.org/abs/2604.06662
