acceptodds
Under review as a conference paper at ICLR 2027

DIVER: Knowledge Theft from GraphRAG System via Structural Diversity-Aware Querying

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) systems organize knowledge as interconnected graphs. While this connected structure facilitates retrieval for reliable reasoning, it also puts the system’s internal knowledge at risk of theft. By continually issuing queries, an attacker can accumulate disclosed entities and relations and progressively recover the underlying knowledge graph. Although existing methods employ various strategies to improve knowledge extraction, identifying promising entities and effectively querying them to recover additional knowledge remains challenging. To this end, we introduce DIVER an efficient GraphRAG extraction framework based on Structural Diversity-Aware Querying. DIVER identifies entities with high remaining extraction potential by combining two complementary signals: (i) local structural diversity, which estimates the potential to reveal structurally distinct knowledge, and (ii) neighborhood query exposure, which accounts for how thoroughly an entity's local region has already been explored. Using these signals, DIVER prioritizes promising entities for targeted relational queries and shifts toward new topics when local extraction opportunities diminish, balancing local exploitation with global exploration. Across Medical, Novel, and Agriculture, DIVER outperforms all evaluated baselines in entity and relation recall and F1 score. These findings highlight the importance of fully exploiting remaining extraction potential for efficient knowledge theft from GraphRAG systems. Code is available at https://anonymous.4open.science/r/DIVER-93FE.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
GraphRAG, GraphRAG Attack

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qamSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes DIVER, a query strategy for extracting knowledge from black-box GraphRAG systems under a limited query budget. The main idea is to avoid repeatedly querying nodes whose neighborhoods contain a lot of overlapping information. DIVER uses BNRR and neighborhood query history to decide which entities to query next. Experiments on three datasets show that DIVER performs better than the baselines considered by the paper in terms of entity and edge recovery.

Critical strengths
  1. The main intuition is clear and reasonable: repeatedly querying highly connected nodes can be inefficient when their neighborhoods contain substantial redundancy.
  2. The method is relatively simple and easy to understand.
Critical weaknesses
  1. The paper discusses several recent graph reconstruction attacks, but does not compare against them experimentally. This makes it difficult to tell how DIVER compares with the strongest existing approaches.
  2. The attacker appears to control an additional instruction given to the response-generation model, rather than only the user's query. The authors should clarify whether DIVER still works when only the user query can be controlled.
Questions for authors

Clarification requests

  1. Why are the graph reconstruction methods discussed in the related work not included in the main comparison?
  2. Does DIVER require control over the response-generation instruction, or can it work when the attacker controls only the user query? Suggestions for future work
  3. Test DIVER on other GraphRAG implementations and LLMs.
  4. Study whether BNRR can reliably predict the amount of new information obtained by a query.
5 comments
AuthorsOpenReview ↗

Response to Reviewer qamS (1/3)

Dear Reviewer qamS,

Thank you for your expertise and thoughtful suggestions. We are glad you found the intuition clear and the method easy to follow. We answer each question below.

W1 & Q1: Regarding the comparison with graph reconstruction attacks in the related work

Thank you for your expertise in this line of work.

We chose baselines according to the object each method recovers. The main comparison already includes AGEA, which matches our budgeted graph-recovery setting most closely, and GRASP, a graph reconstruction attack discussed in our related work, adapted to whole-graph recovery (Table 1, App. B); its precision exceeds 98% on all three datasets while recall stays below 5%. Among the other methods in App. E, RAGCrawler targets coverage of the underlying text corpus rather than the indexed graph. GraphSteal's untargeted method recovers the indexed graph, which makes it a direct comparison.

To complete the picture, we implemented GraphSteal's untargeted method and evaluated it with DIVER and AGEA on G05 under the same index, answering model, Structured Record Response interface and 50-slot budget (mean ± SD over seeds 42–44). The last column counts recovered entities among the 50 most structurally important entities in the index.

MethodNode P (%)Node R (%)Node F1 (%)Edge P (%)Edge R (%)Edge F1 (%)Top-50 entities
DIVER99.33 ± 0.8020.47 ± 2.9433.87 ± 3.9799.59 ± 0.2115.33 ± 2.5926.51 ± 3.8743.3 ± 1.5
GraphSteal99.54 ± 0.1614.64 ± 0.3825.53 ± 0.5997.79 ± 0.9812.16 ± 0.3321.63 ± 0.5334.3 ± 2.3
AGEA99.90 ± 0.0914.32 ± 1.3525.03 ± 2.0598.04 ± 0.5411.77 ± 1.4920.99 ± 2.3935.0 ± 2.0
Paired comparisonSeed 42Seed 43Seed 44
DIVER edge-F1 gain over GraphSteal (pp)+2.09+3.91+8.66

Taken together, the three graph reconstruction attacks show how much the choice of the next entity matters.

  • All methods keep precision near or above 98%, so the differences lie in recall: how much of the graph each method reaches within the budget.
  • GRASP's accurate but narrow recovery shows that precise extraction alone does not produce broad coverage when the order of queries ignores where unseen knowledge remains.
  • GraphSteal's breadth-first expansion and AGEA's degree-weighted selection reach nearly the same recovery on G05, and both recover about 35 of the 50 most structurally important entities. DIVER recovers about 43 of them and has the highest recall and F1, with higher edge F1 than GraphSteal on every paired seed.

Neither traversal order nor observed degree identifies which entities still hold unrecovered relations. DIVER's estimate of remaining potential, which combines neighborhood redundancy with neighbor-propagated query history, is what directs the same budget toward the important parts of the graph.

AuthorsOpenReview ↗

Response to Reviewer qamS (2/3)

W2 & Q2: Regarding attacker control over the user query only

We appreciate your careful attention to the threat model.

DIVER does not require it. DIVER only decides which entity to ask about next and needs no access to the system prompt. In the submission, we separated retrieval and generation inputs so that retrieval is driven solely by the natural-language question our strategy selects, which isolates query allocation from retrieval noise caused by output instructions (Sec. 2.2). This was a choice for controlled evaluation, not a requirement of the method.

To verify this, we kept the operator's default GraphRAG system prompt and sent a single user message to both retrieval and generation, on G05 (mean ± SD over three 50-slot runs):

Single user messageMethodNode R (%)Node F1 (%)Edge R (%)Edge F1 (%)Top-50 entities
Question + record request in the same messageDIVER20.47 ± 2.9433.87 ± 3.9715.33 ± 2.5926.51 ± 3.8743.3 ± 1.5
Question + record request in the same messageAGEA14.32 ± 1.3525.03 ± 2.0511.77 ± 1.4920.99 ± 2.3935.0 ± 2.0
Ordinary question onlyDIVER16.77 ± 1.2527.66 ± 1.786.56 ± 0.4710.74 ± 0.7444.7 ± 0.6
Ordinary question onlyAGEA14.94 ± 0.5225.22 ± 0.635.93 ± 0.209.72 ± 0.2743.3 ± 3.5

These results answer the question directly: DIVER does not need the response-generation instruction.

  • Entities. With ordinary questions alone, DIVER still reaches about 45 of the 50 most structurally important entities, as many as when records are requested. Identifying the important entities of the index therefore requires only the user query.
  • Links. Writing the record request into the same user message, which any user can do, more than doubles edge recall over ordinary questions. The separate instruction channel of the submitted setup is not needed for this level of recovery.
  • Strategy. DIVER exceeds AGEA in recall and F1 under both user messages. The benefit of its query allocation therefore carries over unchanged to the weaker, query-only threat model.

Q3: Regarding other GraphRAG implementations and LLMs

Thank you for this suggestion for broader evaluation.

The submission already varies the target along several axes: three domains with different graph structure (Novel, Medical, Agriculture) and graph-size scaling from G05 to G10 (App. A.6). To extend this to the LLM dimension, we tested DeepSeek-V4-Flash and Qwen3.7-Flash as both answering model and question writer on G05 (Structured Record Response, one 50-slot run per pairing):

Question writerAnswering modelNode P (%)Node R (%)Node F1 (%)Edge P (%)Edge R (%)Edge F1 (%)
DeepSeekDeepSeek99.8917.7830.1999.3713.0923.13
DeepSeekQwen98.2219.8633.0494.4314.6625.38
QwenDeepSeek99.9122.4736.6897.9618.0330.46
QwenQwen99.4220.5034.0090.6615.4326.37

Recovery stays at the same level, and is often higher, when the answering model, the writer, or both come from a different family. DIVER is therefore not tied to one LLM. Because DIVER interacts only through questions and answers and works on the attacker-side recovered graph, it does not depend on the GraphRAG implementation. In line with your suggestion, the revised PDF notes extending the evaluation to other implementations such as LightRAG as future work (App. A.11).

AuthorsOpenReview ↗

Response to Reviewer qamS (3/3)

Q4: Regarding whether BNRR reliably predicts the new information obtained by a query

Thank you for your expertise in this question.

We designed BNRR as a ranking signal for rank admission and validated it that way in the submission. In the controlled branch experiment of App. A.1, three policies start from identical graph states on Novel:

Selection strategyHigh-diversity (BNRR)Degree-onlyLow-diversity
New indexed edges per query25.6923.8119.69
Edge overlap with observed graph (%)30.4636.1337.57

To examine individual queries, we compared pairs of entities at saved graph states that share the same observed degree and prior exposure but differ in pre-query BNRR. Saved states come from three DIVER trajectories per dataset (seeds 42–44), frozen after rounds 20, 40, 60 and 80. Each anchor is evaluated with two independent full writer–GraphRAG queries, and yield counts previously unrecovered directed pairs that match the index (App. A.1, Fig. 5).

DomainMatched pairsNew pairs per query, higher BNRRNew pairs per query, lower BNRR
Novel4216.6211.63
Medical4815.3914.59

Read together, the comparisons show both what BNRR predicts and where it applies.

  • Beyond degree. With degree and query history held fixed, the entity with higher BNRR returns more new relations per query on both Novel and Medical. On Novel the difference is about 43% (16.62 vs. 11.63). BNRR therefore predicts the amount of new information beyond what degree already captures.
  • Where it applies. Since BNRR = √(d·e) with e ∈ [1, d], it always lies in [√d, d], so an entity's BNRR is a lower bound on its degree. The matched pairs in Fig. 5 reach pre-query BNRR of about 19 on Novel and 15 on Medical, so they include entities with large neighborhoods, where redundancy among neighbors can differ substantially. This is the regime BNRR is designed for. With only two or three neighbors, there are at most a few possible links among them, the range [√d, d] is narrow, and BNRR reduces to degree by construction. BNRR therefore adds ranking information once neighborhoods are large enough for redundancy to vary.
  • At the pool level. In the frozen-state experiment, ranking the whole candidate pool by BNRR yields the most new edges and the least overlap, so the per-pair signal also turns into a better allocation of queries.

BNRR is used to rank candidates, and these results show that it ranks reliably in the regime where the ranking matters.

Revisions in the updated PDF (new text in blue):

  • W1 & Q1: App. A.10 adds the comparison with GraphSteal (Table 14, Obs. 18), and App. B describes its implementation. Sec. 4.1 points to this comparison.
  • W2 & Q2: App. A.9 reports the single-user-message and ordinary-question settings (Table 12, Obs. 17). Sec. 2.2 now states that DIVER requires control only over the question.
  • Q3: App. A.11 reports the cross-family victim and query-writer models (Table 15, Obs. 19) and notes the extension to other implementations such as LightRAG as future work.
  • Q4: App. A.1 adds the matched-pair comparison at equal observed degree and exposure (Fig. 5), and Sec. 3.1 points to it.

Thank you again sincerely for your time, efforts, and thoughtful feedback. We have incorporated your helpful comments in the revised manuscript. We would be glad to address any further comments or questions you may have.

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

This paper introduces DIVER, an attack framework to extract proprietary entities and relationships from black-box GraphRAG systems through iterative and targeted querying. They introduce BNRR scoring metric, which scores an entity by combining its observed degree with how sparsely its neighbors are interconnected. Neighborhood query exposure discounts entities whose local region has already been queried. It is along with an adaptive mechanism that dynamically shifts between exploiting known entities and exploring new topics. Experiments conducted across three domain-specific datasets show that it has higher entity and relation recall and F1 than five baselines while keeping precision high.

Critical strengths
  1. Instead of relying on naive node degree for target selection, DIVER introduces the BNRR to balance an entity's neighborhood size against its local structural redundancy. This can efficiently solves the problem of attackers getting trapped in dense, overlapping local clusters.
  2. Do not rely on a single final recall number but check the conclusion from several aspects. They report entity and directed-edge recall, F1, and precision, also covering different structural weightings that agree. Elaluation on efficiency with rounds and tokens. Such results make the claim that structural diversity helps knowledge theft more strong.
Critical weaknesses
  1. The attack relies on a very cooperative victim. The model needs to list every retrieved entity and relation with full descriptions in a format the attacker hands it. The claim is that response-only access can be accumulated into serious knowledge theft, but the setup looks more like a system that hands over its own graph on request. The parser is deterministic and throws away any response it can’t parse into records, so the whole accumulation loop hinges on that compliance.
  2. The baseline that really matters here is AGEA, since it’s the one doing the same thing: black-box graph recovery from structured responses. Against it the margin is thin: +1.05 entity-F1 and +0.54 edge-F1 (Table 1) on Agriculture. Table 4 reports standard deviations for DIVER only, not for AGEA.
Questions for authors
  1. How does DIVER perform against a victim that does not comply with the structured extraction request, e.g., one with a meta-prompt tuned for abstractive summarization.
  2. The victim model (DeepSeek-V4-Flash) is also used as DIVER’s query writer and as the response-to-graph parser. Does this shared-model setup favor the attack, and how does performance change when the victim is a different model family?
  3. TGTB, PIDE, and IKEA have their text responses passed through a separate DeepSeek converter, while DIVER and AGEA read structured responses directly. Can you give all methods the same structured-response interface so the comparison is fair.
6 comments
AuthorsOpenReview ↗

Response to Reviewer LgYn (1/2)

Dear Reviewer LgYn,

Thank you for your expertise and the detailed review. We are glad you found that BNRR keeps the attacker from getting trapped in dense, overlapping clusters, and that our multi-faceted evaluation strengthens the claims. We answer each question below.

W1 & Q1: Regarding victim compliance and an abstractive-summarization meta-prompt

We appreciate your careful attention to the threat model.

We considered this when separating DIVER's components. DIVER's contribution is the query-allocation strategy, which is independent of the elicitation prompt. The structured interface follows the extraction-instruction setting of prior RAG extraction attacks (Zeng et al., 2024; Qi et al., 2025), and DIVER and AGEA share it, so the comparison isolates strategy. The prose conditions below use a shared QA response-to-graph pipeline with an LLM converter instead of the deterministic parser, so no response is discarded for failing to match a record format.

To evaluate a non-compliant victim directly, we used ordinary domain questions on G05, the five-book GraphRAG index in App. A.6 (239 text chunks, 4,994 entities, 7,175 directed edges), under two victim configurations: the default GraphRAG system prompt, and a meta-prompt instructing the system to summarize in concise prose without listing records. Values are mean ± SD over three 50-slot runs. Structured Record Response is included as a reference.

Victim configurationMethodNode R (%)Node F1 (%)Edge R (%)Edge F1 (%)Top-50 entities
Structured Record (reference)DIVER20.47 ± 2.9433.87 ± 3.9715.33 ± 2.5926.51 ± 3.8743.3 ± 1.5
Default QADIVER16.77 ± 1.2527.66 ± 1.786.56 ± 0.4710.74 ± 0.7444.7 ± 0.6
Default QAAGEA14.94 ± 0.5225.22 ± 0.635.93 ± 0.209.72 ± 0.2743.3 ± 3.5
Summarization meta-promptDIVER17.31 ± 0.8828.45 ± 1.216.82 ± 0.1811.11 ± 0.2244.7 ± 2.1
Summarization meta-promptAGEA14.70 ± 1.5524.87 ± 2.306.24 ± 0.7310.27 ± 1.0341.3 ± 1.2
  1. Summarization does not stop leakage. Under the summarization meta-prompt, DIVER still recovers about 45 of the 50 most structurally important entities, as many as with structured records. Its node F1 and edge F1 are no lower than under default QA. Instructing the victim to summarize therefore gives little protection against cumulative extraction.
  2. The accumulation loop does not depend on compliance. Without any structured output, DIVER still builds a growing recovered graph and outperforms AGEA under both victim configurations. Its advantage comes from deciding which entities to ask about, not from the victim's cooperation.
AuthorsOpenReview ↗

Response to Reviewer LgYn (2/2)

Q2: Regarding the shared-model setup and other victim model families

Thank you for this thoughtful question about the experimental design.

We designed the pipeline to keep the parsing step model-free. In the submitted DIVER experiments, the response-to-graph step is a deterministic parser (App. D.3), not DeepSeek. Apart from constructing the fixed index, DeepSeek-V4-Flash appears in DIVER runs only as the victim's response generator and as the question writer. The only possible shared-family effect is therefore between writer and victim.

To test this directly, we varied the victim and the writer independently on G05 under Structured Record Response, using DeepSeek-V4-Flash and Qwen3.7-Flash, the same deterministic parser, and one 50-slot run per pairing (seed 42).

VictimQuestion writerNode P (%)Node R (%)Node F1 (%)Edge P (%)Edge R (%)Edge F1 (%)Top-50 entities
DeepSeekDeepSeek99.8917.7830.1999.3713.0923.1343
DeepSeekQwen99.9122.4736.6897.9618.0330.4645
QwenDeepSeek98.2219.8633.0494.4314.6625.3841
QwenQwen99.4220.5034.0090.6615.4326.3745

The highest recovery comes from a cross-family pairing (Qwen writer against the DeepSeek victim). With the Qwen victim, the same-family and cross-family writers perform comparably. Sharing a model family brings no consistent advantage. DIVER's effectiveness transfers across victim model families and does not rely on writer–victim affinity.

Q3: Regarding the same structured-response interface for all methods

Thank you for this suggestion regarding the comparison protocol.

In the submission, each extraction baseline (TGTB, PIDE, IKEA) used the native output interface from its original paper, so that each method was reproduced faithfully. DIVER and AGEA, both graph-recovery methods, shared the structured interface. To further isolate the effect of query selection, we evaluated DIVER, AGEA and the three extraction baselines on G05 with the same Structured Record Response format, deterministic parser, and 50-slot budget (seed 42).

MethodNode P (%)Node R (%)Node F1 (%)Edge P (%)Edge R (%)Edge F1 (%)
DIVER99.8917.7830.1999.3713.0923.13
AGEA100.0015.8227.3298.5512.3221.90
PIDE99.8613.8024.2497.858.8616.26
IKEA99.6812.6222.4099.688.7716.12
TGTB99.8311.4320.5287.216.7512.52

A shared structured interface brings most baselines to near-perfect precision but does not close the recall gap. DIVER still recovers the largest share of entities and links. The ranking therefore reflects the query-selection strategy, not differences in output format or conversion.

Revisions in the updated PDF (new text in blue):

  • W1 & Q1: App. A.9 reports the default GraphRAG prompt and the summarization meta-prompt (Table 12, Obs. 17). Sec. 2.2 points to these results, and App. B describes the shared QA response-to-graph pipeline.
  • Q2: App. A.11 varies the victim and query-writer models independently (Table 15, Obs. 19), and Sec. 4.1 points to it.
  • Q3: App. A.10 evaluates DIVER, AGEA, PIDE, IKEA and TGTB under the same Structured Record Response interface (Table 13, Obs. 18), and Sec. 4.1 points to it.

We once again sincerely thank you for your time, careful reading, and thoughtful questions, which have helped us further strengthen the paper. We would be glad to address any further comments or questions you may have.

Thank you for the response. Q2 and Q3 are addressed. One concern remains on the non-compliant victim results. I agree that the accumulation loop still runs without structured output. However, the table omits precision, unlike the other two tables. From the reported F1 and recall, edge precision is about 30% under both prose conditions, compared with about 98% under structured records, and edge recall drops by more than half. Does this suggest that relation recovery depends heavily on victim compliance?

Reviewer_E2tMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper observes that the richness of currently observed information does not necessarily imply substantial remaining potential for future knowledge extraction. Based on this insight, the authors propose DIVER, which leverages BNRR and Neighborhood Query Exposure to dynamically allocate between exploration and exploitation and to guide anchor selection. Experiments demonstrate the effectiveness of the proposed method.

Critical strengths

The observation that the richness of currently observed information does not necessarily translate into substantial remaining potential for future knowledge extraction is insightful and practically meaningful.

The experiments in this paper are detailed and thorough.

The paper provides comprehensive details for reproducibility.

Critical weaknesses

Knowledge theft typically involves recovering both specific entities and semantically meaningful relations. In contrast, the paper appears to collapse all relation types into a generic 'related to' relation. This formulation seems closer to structural graph reconstruction, with an emphasis on recovering graph connectivity, rather than on reconstructing knowledge at a more semantically specific level.

Is the methodological novelty sufficient beyond replacing degree-based prioritization with a local structural-diversity heuristic and extending query-count discounting to neighboring nodes?

How does DIVER perform under a standard QA-only victim interface without the explicit instruction to return all retrieved entities, relations, and unsummarized descriptions?

The paper mainly relies on basic statistical visualizations, such as line charts and bar charts. It would be helpful to enrich the visual presentation with more informative or task-specific analyses.

Questions for authors

Refer to the above text.

6 comments
AuthorsOpenReview ↗

Response to Reviewer E2tM (1/3)

Dear Reviewer E2tM,

Thank you for your expertise and your careful reading of our paper. We are glad you found our central observation insightful and our experiments thorough. Below we answer each point, with additional experiments and analyses.

W1: Regarding relation semantics and the related_to formulation

Thank you for this insightful perspective on what counts as knowledge in a GraphRAG index.

We designed the recovery target with this in mind. In Microsoft GraphRAG, each indexed relationship is stored as (source, target, description, weight). The index has no relation-type field: the meaning of each relation lives in its free-text description. Our recovered graph uses the same schema (Sec. 2.2, App. D.3). Every recovered edge keeps its direction and, when the victim provides one, the disclosed description. related_to is only an edge label that the index and the recovered graph share. It does not merge relation semantics. Edge F1 in Table 1 scores endpoints and direction, and the descriptions are kept alongside.

The record below is taken verbatim from one response in the submitted Novel run (seed 42, round 11). It shows what a single recovered edge contains:

FieldValue stored in the recovered graph
Source entityTIBIA UTRICULARIS
Source entity description"The bagpipe, said to be a favourite instrument of Emperor Nero"
Target entityNERO
Edge labelrelated_to (shared with the index)
Edge description"Nero favoured the tibia utricularis (bagpipe)"

To check that this holds across the whole recovered graph, we audited the final DIVER graphs of the main experiments (100 rounds, seeds 42–44). For each recovered edge that matches the index, we measured (i) whether it carries a description and (ii) content-word overlap with the indexed description of the same relation, i.e., the share of content words in the recovered description that also appear in the indexed one. As a reference, we paired each recovered description with a randomly drawn indexed description. Values are mean ± SD over the three runs; the reference is computed on the pooled edges.

DatasetIndexed edgesMatched edges per run (mean)Matched edges with a description (%)Overlap with the indexed description (%)Random-pair reference (%)
Novel29,2362,00699.80 ± 0.1199.97 ± 0.031.3
Medical3,0461,73499.47 ± 0.8299.93 ± 0.125.0
Agriculture9,4111,792100.00 ± 0.00100.00 ± 0.000.9

The audit answers the question of what a recovered relation conveys. Each edge carries a specific predicate, such as who favoured which instrument, not just the fact that two entities co-occur. The gap between the overlap and the random reference shows that this predicate is the one the index stores for that particular pair, not generic text that would fit any edge. In GraphRAG the description is the relation's semantics. An attacker who recovers an edge therefore obtains the indexed statement about how the two entities relate, and Edge F1 counts only the structural part of that recovery.

AuthorsOpenReview ↗

Response to Reviewer E2tM (2/3)

W2: Regarding the methodological novelty beyond degree-based prioritization and query-count discounting

Thank you for your expertise in this question.

Motivation. Under a limited query budget, knowledge theft depends on identifying high-value entities: those whose query is likely to yield as many diverse, previously unseen relations as possible. As you noted in your summary, how much has been observed around an entity does not tell us how much remains to be extracted. Existing strategies cannot tell these apart. AGEA ranks candidates by degree, direct query frequency and recency, and GRASP and GraphSteal expand in FIFO or BFS order. These signals describe how much has been seen around an entity. They do not capture how redundant that neighborhood is, or how much of it earlier queries to neighboring entities have already exposed. As a result, the budget drifts into dense, overlapping regions.

Our design. DIVER addresses this with two signals that have a clear grounding, plus one controller:

  • BNRR: based on the order-two, similarity-sensitive diversity of Leinster and Cobbold, with adjacency treated as similarity. It gives an effective number of non-redundant neighbors, e ∈ [1, d]. BNRR (√(d·e)) balances neighborhood size against redundancy and has a probabilistic interpretation (App. D.1).
  • Neighborhood query exposure: query history propagates to neighbors, because querying a neighbor already discloses edges incident to the entity.
  • Unified controller: BNRR decides rank admission, exposure decides sampling, and their ratio decides between exploration and exploitation (Eqs. 8–9). AGEA instead switches on recent novelty.

Each component accounts for a distinct part of the gain (Table 3; Novel, 100 rounds, seed 42). Relative to the full model (node/edge F1 22.08/15.32), replacing BNRR with degree gives 19.09/13.64, self-only exposure 19.24/13.26, no exposure 18.38/13.11, constant 50% admission 20.36/14.97, and uniform anchor selection with the gate kept 16.38/11.84; exploit-only and explore-only give 17.61/13.11 and 15.79/9.40.

The controlled branch experiment in App. A.1 isolates the selection score itself. Three policies start from the same four frozen graph states, with 16 queries each, and differ only in the admission score:

Selection strategyNew nodes per queryNew edges per queryNode overlap with observed graph (%)Edge overlap with observed graph (%)
High-diversity (BNRR)22.3125.6945.2330.46
Degree-only21.2523.8149.5636.13
Low-diversity16.6319.6954.3837.57

Table 3 and the branch experiment address the two parts of your question separately.

  • Beyond degree. From the same graph states, ranking by BNRR returns a smaller share of already-observed edges and more new edges per query than ranking by degree, and ranking by the inverse of BNRR is worst on every measure. Over a full 100-round run, replacing BNRR with degree lowers both node and edge F1. Neighborhood redundancy is therefore information that degree does not carry, and acting on it changes which queries are worth spending budget on.
  • Beyond per-node query counts. Self-only exposure, which discounts only the queried entity in the manner of a per-node count, recovers less than neighbor-propagated exposure, and removing exposure altogether recovers less still. Queries to neighbors do expose part of an entity's neighborhood, and accounting for this is what keeps the attacker from revisiting regions that are already largely disclosed.
  • The controller as a whole. Keeping the explore–exploit gate but choosing anchors uniformly causes one of the largest drops, comparable to running only one mode. The gain therefore does not come from a single re-weighted score. It comes from using the two signals together for both anchor selection and mode allocation.
AuthorsOpenReview ↗

Response to Reviewer E2tM (3/3)

W3: Regarding the performance under a standard QA-only victim interface

Thank you for this question about realistic deployments.

DIVER contributes the query-allocation strategy (which entity to query and when to explore), which is independent of the prompt used to elicit records. The submission follows the extraction-instruction setting of prior RAG extraction attacks (Zeng et al., 2024; Qi et al., 2025) and gives DIVER and AGEA the identical interface (Sec. 2.2, App. F.5).

To address QA-only access directly, we evaluated three interfaces on G05, the five-book GraphRAG index in App. A.6 (239 text chunks, 4,994 entities, 7,175 directed edges):

  • Structured Record Response: records are requested inside a single user message.
  • Default GraphRAG QA: ordinary domain questions with the default system prompt and no extraction instruction.
  • Summary-Constrained QA: the same questions, with the victim instructed to answer in concise prose.

Each run has 50 query slots. Values are mean ± SD over seeds 42–44. Both methods use the same question writer and response processing. The last column counts recovered entities among the 50 most structurally important entities in the index.

InterfaceMethodNode R (%)Node F1 (%)Edge R (%)Edge F1 (%)Top-50 entities
Structured RecordDIVER20.47 ± 2.9433.87 ± 3.9715.33 ± 2.5926.51 ± 3.8743.3 ± 1.5
Structured RecordAGEA14.32 ± 1.3525.03 ± 2.0511.77 ± 1.4920.99 ± 2.3935.0 ± 2.0
Default QADIVER16.77 ± 1.2527.66 ± 1.786.56 ± 0.4710.74 ± 0.7444.7 ± 0.6
Default QAAGEA14.94 ± 0.5225.22 ± 0.635.93 ± 0.209.72 ± 0.2743.3 ± 3.5
Summary-ConstrainedDIVER17.31 ± 0.8828.45 ± 1.216.82 ± 0.1811.11 ± 0.2244.7 ± 2.1
Summary-ConstrainedAGEA14.70 ± 1.5524.87 ± 2.306.24 ± 0.7310.27 ± 1.0341.3 ± 1.2
  1. The threat does not depend on the instruction. With ordinary questions and no extraction request, DIVER still recovers about 45 of the 50 most structurally important entities, matching the structured setting. Standard QA access is enough to leak the core of the index.
  2. The advantage comes from the strategy. DIVER exceeds AGEA in node and edge recall and F1 under all three interfaces, which confirms that the gain is independent of the elicitation format.

W4: Regarding more informative and task-specific visualizations

Thank you for this constructive suggestion.

Beyond the recovery curves, the submission includes a task-specific controlled branch analysis (App. A.1, Fig. 1) that compares selection policies from identical graph states. To strengthen the visual analysis, we have added three task-specific figures to the revised PDF:

  • (a) Recovered subgraph overlaid on the index (App. A.12, Fig. 10): DIVER recovers 2,132 indexed entities and 2,431 directed pairs versus 1,525 and 1,776 for AGEA, and is less concentrated (12.0% vs. 15.9% of matched entities in the most-represented community).
  • (b) Anchor-selection timeline (App. A.12, Fig. 11): per round, the query mode, the anchor's pre-query BNRR and exposure, and the yield. Exploration is interleaved throughout, and most exploitation anchors have zero exposure when selected.
  • (c) Per-query yield versus BNRR (App. A.1, Fig. 5).

Panel (c) builds on a new matched-pair analysis: with observed degree and prior exposure held fixed, the higher-BNRR entity returns more new indexed directed pairs per query, 16.62 vs. 11.63 on Novel (42 pairs) and 15.39 vs. 14.59 on Medical (48 pairs). This is the per-query mechanism behind the aggregate gains, and Figs. 10–11 show its effect on coverage and along the attack trajectory. We kindly invite you to view these figures in the revised PDF.

Revisions in the updated PDF (new text in blue):

  • W1: App. A.8 (Tables 10–11, Obs. 16); Sec. 2.2.
  • W2: App. A.1 (Fig. 5); Sec. 3.1.
  • W3: App. A.9 (Table 12, Obs. 17); Sec. 2.2; App. B (QA pipeline); App. C.2 (Top-50).
  • W4: App. A.12 (Figs. 10–11, Obs. 20); App. A.1 (Fig. 5).

Thank you again sincerely for your time, efforts, and thoughtful feedback. We have incorporated your helpful comments in the revised PDF. If you have any further comments or suggestions, we would welcome any further feedback.

Comments on the paper

2 comments
Area_Chair_BQCsOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewers,

We are now roughly halfway through the Author–Reviewer Discussion period (which runs through 18 November). The authors have submitted their response and, in some cases, a revised manuscript.

Could you please take a moment to:

  • Read the authors’ rebuttal and any updated version of the paper;
  • Leave a reply to the authors on the points you raised, indicating whether your concerns have been addressed.

Even if your evaluation remains unchanged, a short reply confirming your final position is very helpful, as it allows me to track each reviewer’s stance and write an accurate meta-review.

If you are unable to participate further, please let me know as soon as possible.

Thank you for your time and careful reviewing.

Best regards,

Area Chair, ICLR 2027

AuthorsOpenReview ↗

General Response

Dear Area Chairs,

We sincerely thank you and the reviewers for your time, effort, and valuable feedback during the author–reviewer discussion phase. Below, we briefly summarize the main contributions of our paper and the key points addressed during the rebuttal stage.

Overall Assessment

After the discussion, Reviewers E2tM and qamS raised their ratings from 2 to 3. Reviewer LgYn acknowledged our responses and revisions, and kept their original rating.

Reviewer E2tM: Thank you for your detailed response. I will raise my rating accordingly.

Reviewer qamS: Thanks for the carefully response and especially the additional experiments from authors. […] I am raising my score to 3.

Reviewer LgYn: Thank you for the response. Q2 and Q3 are addressed. […] Thank you for reporting precision and revising Obs. 17.

Reviewers also recognized our central observation as "insightful and practically meaningful" (E2tM), our experiments as "detailed and thorough" (E2tM), BNRR as addressing "the problem of attackers getting trapped in dense, overlapping local clusters" (LgYn), and our main intuition as "clear and reasonable" (qamS).

Summary of Our Work

  • We reveal that the richness of observed information around an entity does not indicate its remaining extraction potential, causing existing strategies to waste query budget on dense, overlapping regions.
  • We propose BNRR, a local structural-diversity signal, and neighborhood query exposure, which propagates query history to neighbors; a unified controller uses both for anchor selection and exploration–exploitation allocation.
  • We conduct comprehensive experiments on three domains and graph-size scaling indexes, where DIVER outperforms all evaluated baselines in entity and relation recall and F1.

Addressed Concerns

  • Threat model (LgYn, qamS, E2tM): DIVER requires control only over the user query. With ordinary questions under the default GraphRAG prompt or a summarization meta-prompt, DIVER still recovers about 45 of the 50 most important entities and outperforms AGEA under every interface (App. A.9).
  • Additional baselines and fair comparison (qamS, LgYn): DIVER outperforms the newly added GraphSteal on every seed, and achieves the highest recall and F1 when all baselines share the identical structured-response interface (App. A.10).
  • Cross-model generality (LgYn, qamS): Across DeepSeek and Qwen as victim and query writer, the best result comes from a cross-family pairing, showing no reliance on a shared model family (App. A.11).
  • Novelty and predictive power of BNRR (E2tM, qamS): From identical graph states, BNRR yields more new edges and less overlap than degree-based selection; with degree and exposure held fixed, higher-BNRR entities return more new relations per query (Table 3, App. A.1).
  • Relation semantics (E2tM): Over 99.5% of recovered edges carry descriptions with 99.9–100% overlap with the indexed relation, conveying specific predicates rather than co-occurrence (App. A.8).
  • Visual analysis (E2tM): We added an index-overlay view, an anchor-selection timeline, and a per-query yield analysis (App. A.12, App. A.1).

Thank you again for your time and expertise throughout this review process. All additional analyses are incorporated into the revised PDF. We hope this summary could facilitate the next stage of review and discussion.

Best regards,

Authors of Submission 23468

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.