acceptodds
Under review as a conference paper at ICLR 2027

Show Me Your Lineage: Query-Only Attribution of Evolving RAG Systems

Abstract

RAG systems rely on evolving external knowledge, making curated knowledge bases valuable but vulnerable to theft, redeployment, and modification. Existing black-box ownership methods verify protected snapshots but do not address attribution after a stolen RAG evolves. We take three steps to bridge this gap. First, we formalize this problem by defining provenance laundering (PL) and its corresponding Provenance Laundering Attack (PLA) game, in which an attacker seeks to evade attribution while preserving task utility and inherited knowledge. We then introduce Update-Equivariant Lineage Coding (UELC). It distributes and refreshes coded epoch evidence across fact preserving carriers, recovers a matching epoch from black-box answers, and uses an authenticated commitment chain to extend the match to its predecessors. Finally, extensive experiments show that UELC retains more attribution evidence than the evaluated baselines under laundering attacks and maintains high true-positive rates throughout authorized evolution. Our work provides a new evolution-aware approach to tracing RAG lineages after theft and subsequent evolution using only black-box access.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
RAG security, retrieval-augmented generation, lineage attribution, provenance laundering, black-box verification, data provenance

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_Y4MDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Update-Equivariant Lineage Coding (UELC), a black-box watermarking framework for provenance attribution in evolving Retrieval-Augmented Generation (RAG) systems. The authors introduce the concept of Provenance Laundering Attacks (PLA), in which an adversary steals a protected RAG knowledge base and subsequently modifies its contents or system components while preserving task utility and inherited knowledge. To address this threat, UELC embeds epoch-specific watermarks into fact-preserving semantic carriers using error-correcting codes, periodically refreshes the watermark during authorized knowledge-base updates, and employs cryptographic commitment chains to authenticate historical lineage relationships. The framework enables knowledge-base owners to recover watermark evidence through black-box queries without accessing the suspect system's internal components. Experiments on FARAD and NQ-50K demonstrate improved robustness over RAG-WM, WARD, and SentinelRAG under several knowledge-base transformations.

Critical strengths
  1. The paper introduces the problem of provenance laundering in evolving RAG systems, extending traditional snapshot-based ownership verification to lineage attribution. This is an interesting research direction for protecting RAG knowledge bases against unauthorized reuse and modification.
  2. The proposed UELC framework combines semantic watermarking, error-correcting codes, authorized watermark refresh, and cryptographic commitment chains to support black-box provenance attribution across multiple knowledge-base versions.
  3. The paper is well-written.
Critical weaknesses
  1. The multi-epoch auditing mechanism has a worst-case query complexity of O(MB)O(MB), raising scalability concerns for continuously evolving RAG systems. The evaluation covers only six epochs, which may not reflect real-world deployments with frequent updates. The authors should evaluate longer evolution histories and explore more efficient epoch identification methods.
  2. The method relies on linguistic characteristics of generated responses for watermark extraction, which may be affected by common RAG deployment practices such as concise answers, fixed output formats, and paraphrasing. In the A6-C experiment, UELC's TPR drops to 0.270 despite retaining 97% of task utility and inherited knowledge, raising concerns about its practical robustness. The authors should evaluate more diverse response formats and clarify the assumptions required for reliable watermark extraction.
  3. In static ownership verification, the proposed method achieves performance comparable to RAG-WM, WARD, and SentinelRAG. Although UELC introduces authorized watermark refresh, existing baselines can also be extended to periodically re-protect updated knowledge bases. The authors consider such rolling baselines in Appendix C.6, where Rolling-WARD achieves a lineage TPR of 0.946 compared with 0.954 for UELC. This relatively small difference raises questions about the extent of the improvement in attribution effectiveness. Moreover, the rolling baselines rely on from-scratch re-protection, whereas UELC uses a specifically designed incremental refresh mechanism. The paper should better isolate the benefits of its individual components and provide a clearer justification for the fairness of the update-cost comparisons.
  4. The experimental evaluation is restricted to two relatively small open-source language models, Qwen3-8B and Llama-3.1-8B. Although black-box access can be simulated using locally deployed models, the paper does not demonstrate the effectiveness of UELC on commercial RAG services powered by closed-source LLM APIs. Since watermark extraction depends on how different generators express retrieved knowledge, its effectiveness may vary significantly across model architectures, scales, and generation policies. Experiments using commercial LLM APIs, such as GPT, Gemini, or Claude, would strengthen the generalizability and practical relevance of the proposed framework.
  5. The paper defines successful provenance laundering under task-utility and inherited-knowledge preservation constraints of Uret≥0.95U_{\mathrm{ret}}\geq0.95 and Cret≥0.90C_{\mathrm{ret}}\geq0.90. However, the practical justification for selecting these thresholds remains unclear. Although the authors evaluate different sanitization intensities, the results suggest that substantial watermark suppression can occur while most of the original knowledge and utility are also preserved. For example, even with 100% canonicalization, the RAG system retains 93.7% task utility and 86.2% inherited knowledge, while the TPR drops further to 0.204, although this setting falls outside the predefined preservation constraints. These results raise concerns about the practical security value of the proposed watermarking mechanism.
Questions for authors

See Weaknesses above.

Reviewer_u5mjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors formalize provenance laundering, a problem where a RAG’s knowledge base can be stolen. They then introduce a game called Provenance Laundering Attack (PLA), wherein an attacker can steal and change the RAG without compromising the RAG’s utility and inherited knowledge. To create a tracer that is robust against such attacks in a query-only setting, the authors propose Update-Equivariant Lineage Coding (UELC), where the code epochs are refreshed and distributed so that the stolen RAGs can be attributed. This proposed framework is tested against existing ones like RAG-WM, WARD, and SentinelRAG and UELC, for instance, received 0.9 TPR in the worst-case across attack families A1-A4 compared to 0.22-0.35 for the baselines. However, UELC’s robustness drops under stronger attacks like A5-S and output canonicalization where it got 0.568 TPR and 0.270 TPR respectively.

Critical strengths

S1. The paper well motivates the problem setup and is comprehensive, w.r.t, providing both a formalization of providence laundering and a formalized attack (PLA) and improved attribution framework (UELC) to address it.

S2. The paper includes a novel setup where the owners would continue to update their RAG system knowledge base and the method still needs to be able to attribute stolen RAGs back to the owner even if PLA happens.

Critical weaknesses

W1. The paper needs significant improvement in writing to provide clarity. There are often times that I have had to rely on the appendix section to understand certain concepts such as the attack families (A1-A6), which should be well-defined in the main paper. There are also terms that I still found to be either overused or underdefined and this includes terms such as ⊥, PRF/KDF, and the soft score L_r. See my questions below.

W2. The main paper also underreports UELC’s weaker results. The strongest attack scenarios, which seems to be A5-S and A6-C, are not well discussed. They are absent in Fig. 3(a), for instance. It is also unclear how often the attacks fall back to the original passage. i.e., the failure rules in the protocols, which makes it hard to understand the true robustness of UELC.

Questions for authors

Questions Q1-Q3 below are clarity questions. They would help me understand the paper a lot better.

Q1. How secret is this secret token (line 070)? If there is a dispute that causes the secret token, salts, and openings to be revealed (line 242), does that make the epoch’s token public and could it weaken other copies? Also, as every k_t seems to be derived from k_master, does one compromise expose all epochs? Q2. I don’t think ⊥ means the same for the one in line 239 and the one in eq. 9 (line 281). Can you differentiate this better? Q3. Why is A5-S missing from Figure 3(a)? Shouldn’t A6-C also be included? The last question here is a suggestion. Q4. Can you expand and cite PRF, BCH, KDF and similar related short terms? This would help improve the readability of the paper.

Reviewer_7HHkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper handles the problem of stolen RAGs by formally defining provenance laundering and provenance laundering attacks. It also introduces UELC (update equivalent lineage coding).

The paper is very hard to follow, with almost no intuition given for the equations and formulation.

The results also need to be analysed a bit more clearly to understand the contributions.

Critical strengths

The paper handles an important problem of updation of stolen RAGs

It formally defines provenance laundering and provenance laundering attacks

It introduces UELC (update equivalent lineage coding)

Critical weaknesses

Fig 1: In one sense, even \tilde{R}_1 and \tilde{R}_2 are derived from R_0 (via \tilde{R}_0). So, why is there no lineage evidence?

Eq (1) is not explained properly and fully. It needs examples.

What is a protected semantic unit F_{i,1}? What is w_{i,u}?

Again, Eq (2) and (3) are just stated, and not explained. There are too many notations used, with almost no attempt to explain them or the intuition behind them.

How does one get or set the values of different \epsilon's in Eq (4)

The results need to be explained more clearly. It is not clear what the exact contributions are.

Questions for authors

The equations, figures, method choices, and results need more clarification to understand the contribution of the work properly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.