acceptodds
Withdrawn from ICLR 2027

Selective Watermarking: Secure Multi-Task Attribution of Language Model Output

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Watermarking for language models answers a binary question – was this text machine-generated? – yet a single model now serves many pipelines and tenants, where the operative questions are which task produced a text and whether auditing one task leaks information about another. We introduce Selective Watermarking, which upgrades detection to secure multi-class attribution: distinct cryptographic keys and heterogeneous embedding channels are routed to disjoint compartments of a single document, and a fail-closed rule attributes a text to a task key or abstains. Our system, SAFIZ, composes an undetectable cryptographic in-sampling core with a zoned lexical channel and a global, invertible orthographic channel, fusing calibrated per-channel -values by Fisher's method. The guarantee is exact: under the pseudorandomness of the keyed function, every compartment's detection score follows a law under any non-matching key, giving a family-wise false-attribution rate of at most with no distributional approximation and no fusion weights to tune. The post-hoc channels supply evidence precisely where no undetectable scheme can embed – short, low-entropy outputs – at the declared cost of system-level undetectability. Across six models, three tasks and documents, SAFIZ attributes with zero misattributions, and null trials stay within the bounds ( family-wise, per key). The cryptographic core leaves text quality untouched: under a third-party model, paired perplexity differences between watermarked and unwatermarked generations is balanced in sign on every model. A controlled base/instruct pair further shows instruction tuning doubling the share of documents on which the undetectable core lacks entropy to embed. Under signal-removal attacks, the system degrades to abstention rather than error: soundness and isolation hold even where recoverability is lost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermarking, large language models, text attribution, model accountability, provenance

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PSLqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Existing language model watermarks only determine whether a text was generated by a model. This paper proposes Selective Watermarking, which determines which task of a single model produced a text. It requires a bounded false-attribution rate, ensures that a verifier holding one task’s key cannot gather evidence from another task’s outputs, and abstains when evidence is insufficient. The authors implement this in SAFIZ, which assigns each task an independent key, embeds a cryptographic watermark during generation and post-hoc synonym substitution in disjoint segments of a document, and applies homoglyph substitution across the whole document. At detection time, SAFIZ combines per-channel p-values with Fisher’s method and either attributes the text to a task or abstains based on a Bonferroni-corrected threshold. The authors prove a bound on the false-attribution rate under two assumptions, pseudorandom function security and channel independence.

Critical strengths
  1. The paper is clearly written and well organized.

  2. The observation that instruction tuning leaves more documents unable to carry a cryptographic watermark is useful.

Critical weaknesses
  1. The guarantee follows mainly from existing statistical methods and cryptographic assumptions, and the technical contribution beyond combining existing components is unclear.

  2. The paper’s setting requires the provider to log a registry for every document, so the provider could instead store the outputs and attribute texts by retrieval, yet the paper does not compare against this approach.

  3. The full system is no more accurate than homoglyph substitution alone, so the experiments do not show the benefit of combining multiple channels.

  4. The paper uses code as its motivating scenario, yet homoglyph substitution generally cannot be applied to code, and no code task is evaluated.

Questions for authors

Please address the concerns raised in the Weaknesses section.

Reviewer_rAUEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes the SAFIZ (Selective Watermark Attribution by Fisher-fused Isolated Zoning) system, which aims to achieve secure multi-task attribution of model outputs. The SAFIZ watermark consists of three complementary channels: a cryptographic channel, an orthographic channel using Unicode confusable characters, and a synonym-based lexical channel. SAFIZ achieves an attribution accuracy of 99.19% on a multi-model test set and can handle low-entropy short texts.

Critical strengths

The motivation that watermarks in practical deployment should be assigned to specific business tasks rather than coarsely tied to the model itself is valid.

Critical weaknesses

The paper is suspected to have been submitted with only AI-generated. The authors did not carefully revise and polish the paper. Here are several pieces of evidence: 1.The paper suffers from disorganized writing logic and is difficult to read. For instance, it contains many overly long sentences with dashes (e.g., line 12, 24 in Abstract). Some sections (e.g., Section 3.3) are split into excessively fragmented subsections. And Section 4.1 crams all content into a single paragraph. 2.From the content of this paper, the authors have conducted numerous experiments. However, there is only one table in the experiment section. In addition, many numerical results are reported within textual descriptions without accompanying figures or tables. 3.Figure 1 is a flowchart with no refinement or polishing. The following points are the other weaknesses of the content: 1.The term “selective” is insufficiently justified. The core mechanism does not introduce a selection operation. This term seems to be misleading. 2.The paper fails to clearly elaborate the differences between the proposed watermarking method and prior watermarking techniques. The work simply assembles classic pre-LLM text watermarking techniques (ASCII/synonym) together with existing LLM watermarking schemes. Moreover, the proposed approach appears to be a straightforward adaptation, shifting the scope of watermark keys from model-specific to task-specific. The authors should provide more explanation to claim the rationale for combining these three watermarking channels within the SAFIZ system.

Questions for authors

N/A. See in Weaknesses

Reviewer_MRVEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces to use different key for watermarking, upgrading detection to multi-class attribution. The proposed method composes an undetectable cryptographic in-sampling core with a zoned lexical channel and a global, invertible orthographic channel, fusing calibrated per-channel p-values by Fisher’s method.

Critical strengths
  1. It proposes a segmentation + role framework for building a segment-wise detectable watermarked text.
  2. It gives formal definitions of soundness and isolation for watermark detection on multiple keys.
Critical weaknesses
  1. Using watermarks generated by different keys is a straightforward idea for multi-class attribution, but the challenges and corresponding contributions are less discussed in this paper.
  2. The paper is motivated by answering the question of "which task produced a given text," but the discussion of multi-bit and multi-user watermarks does not sufficiently distinguish whether they address the same problem.
  3. This paper introduces generation in three stages, but lacks discussion of the rationale and motivation for these design choices.
  4. This paper is hard for humans to read.
Questions for authors
  1. What does this paper mean by "authenticated registry"?
  2. What does this paper mean by "family-wise"?
  3. In Figure 1, when is the content of the post-hoc channels produced? Immediately after s1 is produced? CcryptoC_{crypto} is written as "insampling, zoned by entropy" in line 233, but in Figure 1, the pink square is notated "crypto" and "post-hoc channels".
  4. How about the downstream task performance (ELI5 (qa), XSum (summary)) of text applied with SAFIZ?

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

Related papers

Discussion (0)

Sign in to comment.