acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Based Watermarking for Vector Databases

Abstract

Vector databases are expensive to construct but can be readily replicated, creating a need for effective mechanisms to establish their provenance. Watermarking provides a potential solution; however, conventional coordinate-based watermarks can be disrupted by transformations that preserve Euclidean nearest-neighbor retrieval. To address this limitation, we propose a watermarking scheme that leverages local geometric structure both to identify watermark carriers and to encode the watermark message. In particular, distance ratios and angles provide invariant geometric fingerprints that remain stable under a broad class of transformations. Secret keys are then used to select watermark carriers based on these fingerprints, determine message positions, and select the angles used to encode individual bits. Verification requires only the suspect vector database and the corresponding secret key. We empirically evaluate the proposed scheme on four public datasets, with 50 trials conducted per dataset. Across a range of attacks, including multiple similarity transformations, vector shuffling, and index reconstruction, the method achieves a true-positive rate close to 100% at a 1% false-positive rate. Additional experiments evaluate the scheme's robustness to further forms of data modification and quantify the retrieval-quality loss incurred by attacks designed to compromise the watermark.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermarking, vector database, geometry-based, information retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5juJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript proposes a blind watermarking method for vector databases. The approach leverages local geometric fingerprints to locate carriers and modifies their relative angles for bit encoding. To prevent carriers from being placed in easily detectable regions, the method selects half of the domain region. Fingerprints are constructed by quantizing distance ratios and angles, which are subsequently hashed using a secret key. During embedding, a neighboring element is perturbed to align the corresponding angle with a predefined bit grid. The detection process operates blindly by recomputing the fingerprints and extracting the embedded bits via a majority vote at each bit position.

Critical strengths
  • The proposed method successfully integrates invariant identification, keyed addressing, angular encoding, and blind detection into a single cohesive pipeline. A notable benefit of this approach is that it eliminates the need for original-copy alignment during the extraction phase.
Critical weaknesses
  • A major concern is that the geometric information itself may be altered by the watermark insertion process. The methodology appears to rely heavily on quantization tolerances and data redundancy to withstand this self-induced disruption.

  • The mathematical foundation is sound. However, the hard question is end-to-end stability, and the provided theory stops before this question.

Questions for authors
  1. How can it be verified whether the observed robustness depends primarily on redundancy or on the proposed invariance? Specifically, how much of the performance gain originates from the invariant carrier identification versus the angular encoding?

  2. Since the fingerprint controls essentially every decoding decision, and the embedding process alters these features, how does the system handle boundary effects? Can a small feature change crossing a quantization boundary cause a large, discontinuous change in the keyed assignments?

  3. While the manuscript proves that all message positions are assigned at least one carrier before writing, a core question remains: after the writing process and subsequent attacks, are enough correctly recognized carriers left to successfully recover each bit?

  4. While the experiments are a strength in terms of breadth and transparency, the results appear weaker than what is discussed. How strong is the protection when success is strictly defined by detection at the intended false-positive rate?

2 comments
AuthorsOpenReview ↗

Response to weaknesses/questions 1-3

We thank the reviewer for the thoughtful review and for noting that the method integrates invariant identification, keyed addressing, angular encoding and blind detection into a single pipeline, removing the need for original-copy alignment. We respond to each comment below.

W1 (Self-induced change to fingerprints).

The reviewer is concerned that watermark insertion may alter the geometric information the method relies on. Writing can indeed change a fingerprint: the anchor stays fixed, but rotating one of its neighbours changes the local geometry. Table 3 measures these changes and the bit-reading accuracy after writing. The detector combines votes from several carriers per bit, so individual carriers can fail without losing the bit, as long as correct votes remain the majority. The analyses below quantify this, and they show that the self-induced disruption is well absorbed.

Q1 (Redundancy vs. invariance).

The reviewer asks how much robustness comes from invariant carrier identification, from angular encoding and from redundancy. We isolated each component by changing one part of the full method at a time:

VariantNo attackRotationDeletion 10%8-bit quant.Held-out FPR
Coordinate addressing1000.598880.74
Coordinate writing100197.5860.80
Single vote858524.5270.61
Full method10010098.5920.70

TPR (%) and held-out FPR (%), averaged over SIFT, DEEP, nomic and GIST; 50 positive trials per dataset. Thresholds are calibrated per dataset and variant at a nominal 1% FPR. Single vote uses one randomly selected carrier per bit from the same watermarked database.

Removing either invariant addressing or angular encoding collapses detection under rotation to 0.5–1%, while removing majority voting drops detection to 24.5% under deletion and 27% under quantization. Only the full method performs well in every column. This confirms that each component is necessary and that the gains come from their combination. We will add this ablation to the new version.

Q2 (Boundary effects).

The reviewer asks whether a small feature change crossing a quantization boundary can cause a discontinuous change in keyed assignments. If the fingerprint is unchanged, the key assigns the carrier to the same bit. If it changes, the carrier may be missed or assigned elsewhere. Under the keyed hash, however, such a carrier lands on a pseudo-random position with a pseudo-random target bit, so it acts as a random vote rather than a systematic error. A bit is recovered as long as correct votes outnumber incorrect ones. The ablation above shows how much this redundancy helps: single-vote decoding falls to 24.5% after 10% deletion, whereas the full method retains 94–100% detection (Table 2). Boundary effects are therefore effectively neutralized by design.

W2 & Q3 (Carriers remaining after writing and attacks).

The reviewer notes that Theorem 1 covers carrier assignment before writing, and asks whether enough correct carriers remain afterwards. We measured this directly:

ConditionWritten carriers/bitMatched correct carriers/bitPositions with no correct carrierPositions with correct majorityDetection (%)
After writing; no attack13.7710.310.7061.24100
Rotation13.7710.310.7061.25100
Random deletion 10%13.778.330.7558.5695.5
Sim. transform + 10% del.13.778.310.7758.5496
8-bit quantization13.778.810.8255.9591.5
Random deletion 20%13.776.201.2049.1489.5

Equal-weight averages over the four datasets, 50 trials each; positions out of 64. A matched correct carrier is an originally written carrier recovered with its original fingerprint, position and neighbour pair, and read correctly. Original carrier identities are used only for evaluation, not for detection.

After writing, about 75% of carriers are recovered correctly and over 95% of positions keep a correct majority. Rotation leaves these numbers unchanged. Detection does not require all 64 bits, so it stays high even under 20% deletion. This closes the gap between Theorem 1 and end-to-end recovery and confirms the end-to-end stability of the method. We will add this analysis to the new version.

AuthorsOpenReview ↗

Response to weakness/question 4

Q4 (Detection at the intended FPR).

The reviewer asks how strong the protection is when success is defined strictly as detection at the intended FPR. The paper reports both TPR at a specified FPR and AUC. Under rotation, our method achieves 100% TPR on all four datasets at both 1% and 0.1% FPR, with AUC =1.000= 1.000, whereas TVP's TPR after rotation is 0% on SIFT (Table 10). The table below evaluates frozen thresholds on separate held-out negatives across geometric and classical attacks, including the stronger conditions where detection degrades:

ConditionSIFTDEEPnomicGIST
No attack100100100100
Rotation100100100100
Random insertion 30%10010010098
Gaussian noise, relative scale 0.064%10010010096
Random deletion 10%10010010094
Random deletion 20%84847274
8-bit quantization10010010068
6-bit quantization82122
Held-out FPR (%)0.320.720.880.56

TPR (%) with 50 positive trials per dataset. Thresholds are calibrated on 2,500 scores at a nominal 1% FPR and evaluated on 2,500 separate scores; all rows share the same threshold and negatives.

The achieved FPR is below 1% on every dataset, and detection is at or near 100% under all geometric attacks, insertion, noise, 10% deletion and 8-bit storage. The 6-bit row identifies the minimum workable precision, which gives a practical guideline of 8 bits or more. These results confirm that the method delivers strong protection at the intended FPR. We will add this table to the new version.

We hope these responses fully address the reviewer's concerns, in particular the question of end-to-end stability. The new analyses confirm the robustness of the proposed method, and we would be grateful if the reviewer would consider raising the score.

Reviewer_PP4SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a geometry-based watermarking method for vector databases that aims to preserve retrieval performance while remaining robust to geometric transformations. Rather than relying on absolute coordinates, the method uses local geometric structures to identify watermark carriers and encodes watermark information by modifying angles between vectors.

Critical strengths
  1. Comprehensive experimental evaluation. The experiments cover four datasets and a wide range of attacks, including geometric transformations, deletion, insertion, noise, quantization, index rebuilding, and targeted attacks.
  2. Clear presentation. The paper is well-written, clearly structured, and easy to follow.
Critical weaknesses
  1. Limited significance of the outlier deletion experiment. The reported 100% detection rate after removing 40% of outliers is not particularly convincing, since the watermarks are embedded in the 50% of vectors closest to the centroid. Removing outliers may therefore leave most watermark carriers unaffected.
  2. Potential bias in FPR evaluation. The same 5,000 unwatermarked scores appear to be used for both threshold calibration and FPR evaluation. Using separate sets for calibration and evaluation would provide a more reliable estimate of the actual FPR.
  3. Limited robustness under strong attacks. The TPR drops to 30%–50% after 20% random deletion and further decreases to only 2%–12% under 6-bit quantization. These results raise concerns about the method's robustness in challenging scenarios.
Questions for authors
  1. Unit-norm preservation. The proposed angle-based encoding rotates neighboring vectors around a selected anchor. If the original vectors are normalized to lie on the unit hypersphere, does the embedding process strictly preserve their unit norms?
  2. Stability of geometric fingerprints. Since watermark embedding modifies the local geometry of vectors, could it also change the geometric fingerprints used to identify watermark carriers? If so, how does the method ensure reliable carrier identification after embedding, and could this issue be mitigated?
1 comment
AuthorsOpenReview ↗

Response to weaknesses/questions

We thank the reviewer for the positive assessment of our comprehensive experimental evaluation and clear presentation. We respond to each comment below.

W1 (Outlier deletion).

The reviewer notes that outlier deletion may leave carriers unaffected, since they lie in the inner half of the database. We agree, and this is why the paper includes it only as a database-cleaning check. The paper also reports random deletion, which removes carriers directly. After randomly deleting 10% of vectors, detection remains at 94–100% across all four datasets at a nominal 1% FPR (Table 2). The watermark therefore survives both cleaning and direct loss of carriers.

W2 (Bias in FPR evaluation).

The reviewer points out that the same unwatermarked scores appear in both calibration and evaluation. We thank the reviewer for this point and have separated the two. Thresholds are now calibrated on 2,500 unwatermarked scores at a nominal 1% FPR, frozen, and evaluated on 2,500 disjoint scores:

DatasetCalibration: false positivesCalibration: rate (%)Held-out evaluation: false positivesHeld-out evaluation: rate (%)
SIFT1/25000.044/25000.16
DEEP2/25000.086/25000.24
nomic2/25000.0810/25000.40
GIST2/25000.088/25000.32

Each split contains 2,500 scores (50 databases ×\times 50 keys), with different database subsamples from the same 60,000-vector pool.

On unseen negatives, the achieved FPR stays well below the 1% target on every dataset. This confirms that the calibration generalizes and that the reported detection rates are not inflated. We will add this table to the new version.

W3 (Strong attacks).

The reviewer is concerned that TPR drops to 30–50% after 20% random deletion and to 2–12% under 6-bit quantization. Both attacks damage the database far more than they damage the watermark. Deleting 20% of the vectors reduces recall@10 by 20.0–20.4 points, so the attacker sacrifices a fifth of the database's utility, whereas at 10% deletion detection remains at 94–100% (Table 2). The 6-bit experiment was designed to find the minimum workable storage precision. At 8 bits, detection remains at 100% on SIFT, DEEP and nomic (Tables 11–14), which gives practitioners a clear guideline of 8 bits or more. Most importantly, under a rotation that leaves retrieval unchanged, TVP, RS and SCPW fall to random-guessing accuracy and TabularMark's recovery metric falls to zero, while our method retains 100% detection on all four datasets (Tables 1 and 10). This confirms the clear advantage of the proposed method.

Q1 (Unit-norm preservation).

The reviewer asks whether embedding strictly preserves unit norms. It does not: rotating a neighbour about its anchor preserves its distance to the anchor, but not necessarily its distance to the origin. Euclidean search does not require unit norms, however; what matters is whether the correct neighbours are retrieved. We measure this directly with recall@10 (Table 7), and the loss is small (0.3–1.85 points). Retrieval quality is therefore well preserved.

Q2 (Stability of geometric fingerprints).

The reviewer asks whether embedding changes the fingerprints, and how carriers are still identified reliably. Writing can change some fingerprints: the anchor stays fixed, but rotating one of its neighbours changes the local geometry. Table 3 quantifies these changes and the bit-reading accuracy after writing. Because each bit is carried by several carriers and decoded by majority vote, losing some carriers does not lose the bit. The end-to-end effect is reported in Tables 11–14 in the appendix, which we understand the reviewer was not required to read; we therefore highlight it here. These tables show 100% detection on all four datasets, both without attack and under the tested similarity transforms, at a nominal 1% FPR. All of these results are measured after writing, so they already include any fingerprint changes caused by embedding. The method's reliability after embedding is therefore confirmed directly.

We hope these responses fully address the reviewer's concerns. The results confirm that the method is accurate, well calibrated and robust, and we would be grateful if the reviewer would consider raising the score.

Reviewer_oKbSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a blind watermarking scheme for vector databases built from two components: geometric fingerprints of quantized nearest-neighbour distance ratios and angles, which are invariant under similarity transforms and let keyed hashing identify and address carriers; and quantization-index modulation, which writes each bit into a local angle and recovers bits by majority vote, so that detection requires only the suspect database and the secret key. Its main weakness is that the robustness at its core, survival of similarity transforms, protects against a scenario that plausibly never arises in practice, while routine operations can remove the watermark more cheaply than the attacks it defends against.

Critical strengths
  • The paper identifies a real gap in prior work by showing that all four coordinate-based baselines break under similarity transforms.
  • The method has the lowest embedding cost in the comparison and is the only one that survives 8-bit quantization and small Gaussian noise.
Critical weaknesses
  1. The similarity-transform scenario does not survive scrutiny of how evidence would be produced. Detection on a transformed database presupposes that the suspect can submit one. Let R be the transform (rotation, reflection, permutation, or sign flip). Queries come from the original embedding model, so a transformed database can answer queries only if the suspect's pipeline applies R to every query; an in-use database is therefore either untransformed or bound to R. If the suspect submits the database together with R, the inverse transform restores the original frame and detection proceeds on an undistorted database. If the suspect submits only the transformed database, it cannot answer queries from the model the suspect claims to use, and a single probe query from the owner's model demonstrates this. Such a database is evidently not the one actually in use, and the one actually in use, which is effectively undistorted, must be submitted instead. Every branch converges to an undistorted database, so the situation the scheme defends against does not arise. Please specify the forensic setting in which detection must run on a transformed database while R is unavailable.

  2. The removal-cost claim is contradicted by the paper's own data. Section 5.2 and Table 5 conclude that suppressing detection costs 8.6 to 20.2 recall@10 points under the algorithm-aware angle attack. Tables 11 to 14 show that 6-bit quantization drives TPR at 1% FPR to 0.12 or below on all four datasets at a cost of 4.5 to 10.5 points. This cost is below the angle attack on every dataset, and quantization is a routine storage operation that requires no knowledge of the watermark. The cheapest removal path in the paper's own evaluation therefore costs around 5 points, not 8.6. Please reconcile the removal-cost conclusion with Tables 11 to 14, and report the cheapest removal path across all evaluated attack types.

  3. Dilution is untested at realistic ratios. The selection region is the inner half of the database by distance to its centroid, so it must be recomputed whenever the corpus composition changes. Merging stolen vectors into a larger corpus of one's own is a natural use of stolen data, and it directly perturbs the centroid, the region boundary, and the neighbourhoods that fingerprints depend on. The evaluation covers insertion at 30% only, and ratios of 1:1 and above are absent. Please evaluate detection after dilution at 1:1 and higher ratios.

  4. The contribution of embedding over pure fingerprint matching is not established. Detection works by recomputing geometric fingerprints in the suspect database and re-identifying carriers, which presupposes that fingerprints survive attacks on their own. The same capability supports a scheme with no embedding at all: match fingerprints between the owner's original and the suspect database, requiring that one single orthogonal transform maps every matched vector in the original to its matched vector in the suspect database. Hundreds of matches under one transform would prove that the suspect database derives from the original, at zero retrieval loss. Embedding instead costs 0.3 to 1.85 recall@10 points and perturbs the fingerprints of 28 to 45% of anchors (Table 3). Please provide a controlled comparison against this no-embedding variant so that the contribution of the embedding step is measurable.

Questions for authors
  1. What is a realistic forensic setting in which the detector obtains the suspect's vector data but not the query pipeline, given that a transformed database can answer queries only if the pipeline applies the transformation R to every query?

  2. Tables 11 to 14 show that 6-bit quantization drives TPR at 1% FPR to 0.12 or below on all four datasets at a cost of 4.5 to 10.5 recall@10 points, below the angle attack on every dataset. Given this, does the removal-cost conclusion of 8.6 to 20.2 points in Section 5.2 still stand, and what is the cheapest removal path across all attacks you evaluate?

  3. How does detection perform when the stolen corpus is merged into a corpus of equal or larger size before detection, given that the selection region, the centroid, and the fingerprints all depend on corpus composition? The evaluation covers insertion at 30% only.

  4. What does the embedding step contribute beyond a no-embedding variant that matches fingerprints between the owner's original and the suspect database under one global orthogonal transform, and has this comparison been measured?

2 comments
AuthorsOpenReview ↗

Response to weaknesses/questions 1-3

We thank the reviewer for the careful reading and for recognizing that our work identifies a real gap: all four coordinate-based baselines break under similarity transforms, while our method has the lowest embedding cost and is the only one that survives 8-bit quantization and small Gaussian noise. We respond to each comment below.

W1 & Q1 (Forensic setting).

The reviewer argues that a transformed database can answer queries only if the suspect applies RR to every query, so a probe query from the owner's model would always force an undistorted database. We respectfully disagree, because this assumes the suspect uses the owner's embedding model. The suspect can instead deploy f′=R∘ff' = R \circ f and present it as its own model. Queries are then embedded directly in the rotated frame, and the database returns exact Euclidean results, since ∥Rq−Rxi∥2=∥q−xi∥2\lVert Rq - Rx_i \rVert_2 = \lVert q - x_i \rVert_2. The owner's probe fails because the suspect appears to use a different model, not because the database is unused. The suspect never has to disclose RR, and it has every incentive not to. The same setting arises for redistributed copies (dataset hubs, data marketplaces, leaked dumps) that have no query pipeline at all. In this setting, TVP, RS and SCPW recover bits at random-guessing accuracy and TabularMark's recovery metric falls to zero (Table 1), while our method detects 100% of watermarked databases on all four datasets (Table 10). This is precisely the gap our method closes.

W2 & Q2 (Removal cost and quantization).

The reviewer notes that 6-bit quantization suppresses detection at a lower recall cost than the angle attack in Sec. 5.2 and Table 5. We clarify that the 8.6–20.2-point range in Table 5 refers to the whole-database angle attack only, not to a minimum across attack families. The new comparison below reports the lowest-cost condition across all evaluated attacks:

DatasetLowest-cost condition reaching AUC ≤0.6\le 0.6Recall@10 loss (pp)
SIFTSelection-region angle, ρ=0.50\rho = 0.506.31
DEEP6-bit quantization5.64
nomicBoundary, β=0.30\beta = 0.30, m=3m = 34.06
GISTBoundary, β=0.30\beta = 0.30, m=3m = 38.77

Loss relative to the original unwatermarked database. Δθ=0.12\Delta\theta = 0.12; β\beta: displacement budget relative to the median nearest-neighbour distance; mm: fingerprint features targeted per centre. 50 positive trials per dataset; thresholds calibrated on 2,500 scores (nominal 1% FPR) and evaluated on 2,500 separate scores.

Importantly, the 6-bit experiment was included to identify the minimum storage precision at which the method works reliably, not as a routine operation. The answer is clear and practically useful: at 8 bits, detection remains at 100% on SIFT, DEEP and nomic (Tables 11–14). Practitioners therefore have a concrete guideline, namely to store watermarked vectors at 8 bits or more. We view this as practical guidance rather than a negative result. For comparison, a simple rotation removes every coordinate-based baseline at zero retrieval cost. Our method raises the cheapest removal cost to 4.1–8.8 recall@10 points, which confirms a substantial gain in robustness. We will add this table to the new version.

W3 & Q3 (Dilution at 1:1 and above).

The reviewer asks whether detection survives merging the stolen corpus into one of equal or larger size, since the centroid, selection region and fingerprints depend on corpus composition. We ran this experiment. Real, non-overlapping vectors from the same dataset were added, and the centroid, selection region, nearest neighbours and fingerprints were all recomputed on the merged corpus.

True-positive rate (%)

AddedTotal vectorsSIFTDEEPnomicGIST
0%42,000100100100100
30%54,60010010010098
50%63,00098929390
70%71,40093919286
100%84,00089879285

Recall@10 loss due to the watermark (pp)

AddedTotal vectorsSIFTDEEPnomicGIST
0%42,0000.320.400.591.85
30%54,6000.280.360.511.58
50%63,0000.250.330.481.42
70%71,4000.230.310.441.31
100%84,0000.200.280.391.15

Added percentages are relative to the original 42,000 vectors. 50 positive trials per dataset and ratio. Loss compares watermarked and unwatermarked mixtures with identical added vectors and queries.

With thresholds fixed after calibration at a nominal 1% FPR, the held-out FPR on unwatermarked mixtures of each size stays between 0.44% and 1.00%. Even when the stolen data make up only half of the merged corpus, detection remains at 85–92%, and the retrieval cost of the watermark stays below 2 points. This confirms that the method remains reliable under realistic dilution. We will add this table to the new version.

AuthorsOpenReview ↗

Response to weakness/question 4

W4 & Q4 (Embedding vs. fingerprint matching).

The reviewer suggests that matching fingerprints between the original and the suspect database under one global orthogonal transform could prove derivation without embedding. We thank the reviewer for this suggestion, but the matcher answers a different and weaker question. Fingerprints are properties of the data, so matching shows only that two databases share the same source vectors. When those vectors come from a public corpus and a public model (the common case), an independent party produces identical fingerprints legitimately, and the matcher cannot separate copying from independent re-embedding. A keyed watermark proves possession of the owner's marked copy. Moreover, the matcher requires the original vectors and reference fingerprints at verification, and it assumes a single exact global transform. An attacker may instead apply nonuniform local angular changes followed by deletion, insertion or quantization, which no single orthogonal map describes. Our method is blind: from the secret key and the suspect database alone, the detector reconstructs carrier addresses and decodes the key-derived message. This blind, key-based verification is the core contribution, and it comes at a small embedding cost (0.3–1.85 recall@10 points).

We hope these responses fully address the reviewer's concerns. The new experiments confirm the strength of the proposed method, and we would be grateful if the reviewer would consider raising the score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.