acceptodds
Under review as a conference paper at ICLR 2027

Hardening GraphRAG Against Poisoning Across the Knowledge Lifecycle

Abstract

Graph retrieval-augmented generation (GraphRAG) constructs knowledge graphs (KGs) from external corpora to provide structured knowledge for large language models (LLMs), improving factual accuracy and knowledge utilization. However, recent studies have shown that GraphRAG is vulnerable to poisoning attacks, where attackers inject malicious texts into the external corpus to contaminate the KG and manipulate responses to target queries. Existing defenses are largely designed for conventional RAG or focus on repairing already-poisoned KG, leaving the broader poisoning process insufficiently protected. In this paper, we propose SafeG, a novel poisoning defense framework for GraphRAG that spans both KG evolution and KG-based inference. During KG evolution, SafeG quantifies the structural deviation induced by each incoming document through changes in KG information propagation and filters anomalous content. During inference, SafeG constructs multi-granularity evidence views under different retrieval budgets and generates candidate responses, followed by LLM-based credibility verification for robust response selection. Extensive experiments across multiple datasets, frameworks, and LLMs demonstrate that SafeG consistently improves accuracy while reducing attack success rates against GraphRAG poisoning. Overall, SafeG provides an effective lifecycle-oriented defense against structural contamination in GraphRAG.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Graph-based Retrieval-Augmented Generation, Large Language Models, AI Security

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_nnvXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes SAFEG, a two-stage defense for GraphRAG poisoning across the knowledge lifecycle. During KG evolution, each incoming document is hypothetically integrated into the existing graph and scored by the Jensen–Shannon divergence between random-walk stationary distributions before and after insertion. Documents whose structural divergence exceeds a validation-selected threshold are filtered. During inference, SAFEG retrieves nested subgraphs and generates a candidate answer from each evidence view, constructs a retrieval-independent knowledge prior from the LLM’s parametric knowledge, and uses an LLM verifier to select the candidate most consistent with that prior.

Critical strengths
  1. The lifecycle framing is well motivated and addresses two distinct failure points.
  2. SAFEG is evaluated across multiple datasets, GraphRAG frameworks, LLMs, attacks, and a comparatively large baseline set. The authors also include threshold sensitivity, component ablations, cross-attack tests, and a second LLM-judge agreement check.
Critical weaknesses
  1. The evolution-time detector relies on a fragile assumption that poisoned documents induce unusually large global structural change. The core decision rule treats large stationary-distribution divergence as suspicious. However, an adaptive attacker can plausibly distribute a false claim over low-impact entities/relations, mimic the structural footprint of benign updates, or split one attack across multiple small documents. Conversely, benign but important updates can legitimately introduce new hubs or strongly alter graph connectivity. The current attacks are not adapted to the proposed divergence statistic.
  2. The inference-time verifier assumes that the LLM’s parametric knowledge is a trustworthy reference, which conflicts with a core motivation for RAG. The verifier favors answers consistent with an internally generated knowledge prior. This can help when poisoning contradicts well-known facts, but it is potentially harmful exactly when retrieval is needed most: recent events, long-tail facts, domain-specific knowledge, or information absent/incorrect in the base model.
Questions for authors
  1. Resolve the issues mentioned in the Weakness
  2. Can the authors evaluate Stage II on facts that are deliberately absent from or incorrect in the verifier model’s parametric knowledge—for example, post-training temporal facts or private facts that are present only in the clean KG?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/2)

Thank you very much for your careful and constructive review. We greatly appreciate your recognition that "the lifecycle framing is well motivated and addresses two distinct failure points" and that "SAFEG is evaluated across multiple datasets, GraphRAG frameworks, LLMs, attacks, and a comparatively large baseline set." We also greatly appreciate your insightful suggestions for further strengthening our work. We will address your comments and questions individually in the following five parts:

R1: Structural Deviation as a Poisoning Indicator (W1/Q1)

Thank you for your comment. Stage I is motivated by the observation that poisoning attacks often introduce abnormal entity-relation associations, leading to measurable changes in global KG information propagation. By analyzing shifts in the stationary distribution, SafeG captures the collective structural impact of incoming knowledge updates rather than simply counting newly introduced entities or edges.

As shown in Figure 10, poisoned documents exhibit larger structural deviations than benign documents across four datasets, achieving AUROC values of 0.847, 0.769, 0.814, and 0.797, respectively. These results demonstrate the effectiveness of structural deviation as a discriminative signal for identifying suspicious knowledge updates. Stage II further complements structural screening by mitigating residual poisoning during answer generation.

R2: Robustness Against Low-Deviation Adaptive Attacks (W1/Q1)

Thank you for your suggestion. To evaluate whether adaptive attackers can exploit low structural deviations to bypass Stage I, we conducted additional experiments considering two strategies: (1) Relation Rewiring, which manipulates relationships among potentially existing entities to limit structural changes, and (2) Split-Payload Poisoning, which distributes malicious knowledge across multiple documents to reduce per-document structural deviations. Both attacks operate under the original black-box document-injection threat model.

We evaluated both strategies alongside Regular Poisoning on Graph-Novel and HotpotQA using LightRAG with Llama-3.1-8B, with 50 attack attempts per strategy and dataset. We report the Stage I pass rate, defined as the proportion of attack attempts in which all injected documents pass structural screening, together with the ASR of Vanilla GraphRAG and SafeG.

DatasetAttackStage I Pass ↑Vanilla ASR ↑SafeG ASR ↓
Graph-NovelRegular24.0046.0012.00
Graph-NovelRelation Rewiring66.0024.0012.00
Graph-NovelSplit-Payload64.0032.0014.00
HotpotQARegular28.0044.0010.00
HotpotQARelation Rewiring68.0026.0012.00
HotpotQASplit-Payload68.0028.0016.00
AverageRegular26.0045.0011.00
AverageRelation Rewiring67.0025.0012.00
AverageSplit-Payload66.0030.0015.00

The results reveal a clear trade-off between structural evasion and attack effectiveness. Compared with Regular Poisoning, Relation Rewiring and Split-Payload Poisoning increase the average Stage I pass rate from 26.00% to 67.00% and 66.00%, respectively, while reducing Vanilla ASR from 45.00% to 25.00% and 30.00%. This indicates that reducing structural deviations improves screening evasion but does not necessarily strengthen end-to-end poisoning. Furthermore, SafeG reduces the ASR of Relation Rewiring and Split-Payload Poisoning to 12.00% and 15.00%, respectively, demonstrating its effectiveness against both adaptive strategies. We have incorporated these additional experiments and analyses into Appendix E .

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/2)

R3: False Rejections of Benign Knowledge Updates (W1/Q1)

Thank you for highlighting this concern. Figure 11 examines the trade-off between poisoning detection recall and benign false-positive rate (FPR), with the detection threshold selected to balance security and knowledge preservation. To further evaluate the impact on legitimate knowledge evolution, we conducted an additional analysis focusing on high-impact benign updates, defined as the top 20% of benign documents exhibiting the largest structural changes.

DatasetOverall Benign FPR ↓High-impact Benign FPR ↓High-impact Acceptance ↑
Graph-Medical18.50%25.00%75.00%
Graph-Novel23.00%32.50%67.50%
HotpotQA20.50%27.50%72.50%
2WikiMultihopQA21.50%30.00%70.00%
Average20.88%28.75%71.25%

The results show that Stage I preserves the majority of legitimate knowledge updates even when they introduce substantial structural changes. Although high-impact benign updates exhibit a higher FPR than the overall benign population (28.75% vs. 20.88%), 71.25% of these updates are still accepted. These findings demonstrate that substantial structural changes do not necessarily trigger rejection and quantify the trade-off between poisoning detection and legitimate knowledge preservation.

R4: Reliability of Parametric Knowledge in Stage II (W2/Q1)

Thank you for your comment. Stage II generates candidate answers from retrieved KG evidence and uses independently generated parametric knowledge as an auxiliary reference for credibility verification. This design allows the verifier to assess candidate reliability while retaining external knowledge as the basis for answer generation.

To evaluate its effectiveness when the LLM's internal knowledge is insufficient, we conducted additional experiments using LightRAG with Llama-3.1-8B across four datasets. Specifically, we selected queries where No-GraphRAG answers incorrectly but clean GraphRAG answers correctly, representing cases where external knowledge is essential for accurate responses.

Dataset# QueriesClean GraphRAG ACC ↑+ Stage II ACC ↑
Graph-Medical32100.00%84.38%
Graph-Novel36100.00%83.33%
HotpotQA46100.00%84.78%
2WikiMultihopQA43100.00%81.40%
Overall157100.00%83.44%

The results show that Stage II preserves 131 of 157 correct KG-supported answers (83.44%) on queries where the base LLM fails without retrieval. This demonstrates that credibility verification can retain most valid external knowledge even when the model's parametric knowledge is insufficient.

R5: Performance under Unknown or Incorrect Parametric Knowledge (W2/Q2)

Thank you for your suggestion. To directly evaluate Stage II when the verifier's internal knowledge is absent or incorrect, we further analyzed the knowledge references generated for 68 queries from Graph-Medical and Graph-Novel.

We classify these queries into two categories: (1) Unknown Prior, where the generated knowledge reference lacks the facts required to answer the query, and (2) Incorrect Prior, where the reference contains information inconsistent with the ground truth. We then evaluate whether Stage II preserves the correct answers supported by the clean KG.

DatasetPrior Condition# QueriesStage II CorrectStage II ACC ↑
Graph-MedicalUnknown Prior181688.89%
Graph-MedicalIncorrect Prior141178.57%
Graph-NovelUnknown Prior211780.95%
Graph-NovelIncorrect Prior151386.67%
OverallUnknown Prior393384.62%
OverallIncorrect Prior292482.76%

Stage II preserves 84.62% and 82.76% of correct KG-supported answers under unknown and incorrect parametric knowledge, respectively. These results demonstrate that Stage II can retain most valid external knowledge even when the generated reference is incomplete or inaccurate, while also quantifying the remaining influence of unreliable priors.

We sincerely thank you again for the careful evaluation and constructive suggestions. We believe these additional experiments and clarifications have strengthened the manuscript.

Thanks for your rebuttal. it solve most of my concerns. However, I still suggest that the author should further enhance the theoretical analysis. I will raise my score to 3

We sincerely thank you for your recognition of our rebuttal and for raising your score to 3. We greatly appreciate your constructive feedback and will further strengthen the theoretical analysis in the revised manuscript. Thank you again for your valuable time and support.

Reviewer_kUAoOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a two-stage defense against poisoning in GraphRAG systems. At knowledge ingestion, Stage I measures the Jensen–Shannon divergence between random-walk stationary distributions before and after a document-induced graph update, rejecting updates with sufficiently large deviations. At answer generation, Stage II constructs candidate answers from nested subgraphs with different retrieval budgets and uses the model’s internal knowledge as a reference for credibility assessment and selection. The evaluation spans multiple GraphRAG frameworks, datasets, models, and existing attacks.

Critical strengths

The placement of the defense is well motivated. Poisoning may first alter a persistent knowledge graph and then affect many subsequent queries. Intervening at both ingestion and answer generation addresses this lifecycle more directly than filtering only the final retrieved text.

The ablations support an incremental contribution from each stage. In Table 1, the complete system obtains average ACC/ASR of 59.00/15.42. Removing ingestion defense changes these values to 39.83/41.00, while removing inference enhancement gives 46.25/32.59 . These results provide meaningful evidence that the two components are not redundant.

The appendices contain useful diagnostic evidence rather than only additional headline results. They report benign-versus-poisoned divergence distributions, AUROC and threshold analyses, cross-model verifiers, and runtime costs. For example, Stage I adds approximately 0.45 seconds per document, while Stage II increases query time from 2.12 to 5.43 seconds (Appendix E). These analyses help characterize both effectiveness and deployment cost.

Critical weaknesses

W1. Structural distribution shift is not equivalent to semantic poisoning.

Stage I depends on the information encoded in the transition matrix A. If an update changes relation semantics while preserving the node set and A, the stationary distribution remains unchanged and the JS divergence is zero. Conversely, a legitimate update containing new entities or relations may produce a substantial structural change (Equations (3)–(7)). This is a conditional observation about the mechanism, not a demonstrated bypass of the implementation. Whether it is relevant depends on precisely how relation labels, edge weights, and multiple edges enter A. The existing length-balanced comparisons and AUROC results show that the score contains useful signal for the evaluated attack distribution. They do not yet establish whether that signal identifies harmful content rather than correlated structural changes, such as the number of newly introduced nodes or edges.

A focused control contrasting structurally similar false updates with structurally novel but correct updates would substantially clarify what Stage I detects. Reporting false-rejection rates on the latter is particularly important for a system intended to support continuing knowledge updates.

W2. The boundary for cumulative, individually low-divergence updates is unclear.

In Equation (7), documents within a batch are screened against the same previous graph, G^(t−1). Multiple updates that each fall below the threshold could, in combination, induce a larger change or accumulate harmful effects. The experiments with different attacks and poisoning amounts are useful, but they do not directly resolve this potential composition issue (Equation (7); Algorithm 1). I am not asking for a new attack paper. A bounded, defense-aware stress test or a controlled sequence of individually low-divergence updates would help determine whether the issue materially affects final ACC and ASR. Such a test should preserve the paper’s black-box attacker assumptions. Giving the attacker additional access to internal graph state would change the threat model and should not be treated as a directly comparable evaluation.

Questions for authors
  1. How exactly are relation labels, edge weights, and multiple edges represented in A? What are the detection and false-rejection rates for structurally similar false updates and structurally novel correct updates?
  2. Can several individually low-divergence updates, within one batch or across multiple rounds, accumulate harmful effects despite document-level screening? What does a bounded stress test under the original black-box assumptions show?
  3. For trusted facts that are unknown to the model or conflict with its prior, how much does Stage II change answer accuracy?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/2)

Thank you very much for your careful and constructive review. We greatly appreciate your recognition that "the placement of the defense is well motivated," that "the ablations support an incremental contribution from each stage," and that "the appendices contain useful diagnostic evidence" regarding effectiveness and deployment cost. We also greatly appreciate your insightful suggestions for further strengthening our work. We will address your comments and questions individually in the following six parts:

R1: Transition Matrix Construction (W1/Q1)

Thank you for your question. In Stage I, the transition matrix is constructed by row-normalizing the weighted adjacency matrix. Edge weights default to 1, and parallel edges are aggregated by summing their weights when a multigraph is used. Relation labels are stored as edge attributes but do not directly contribute to transition probabilities.

Accordingly, Stage I captures changes in global information propagation rather than directly verifying relation semantics. Semantic modifications that leave the transition matrix unchanged do not affect the stationary distribution. Nevertheless, the structural signal provides effective discrimination in our evaluated settings, achieving AUROC values of 0.847, 0.769, 0.814, and 0.797 across four datasets. We have clarified the transition matrix construction and its detection scope in the revised manuscript.

We construct the transition matrix by row-normalizing the weighted adjacency matrix of the KG. Each edge is assigned a default weight of 1, and parallel edges are aggregated by summing their weights in multigraph representations. Relation labels are retained as edge attributes, while transition probabilities are determined by the normalized edge weights. This construction captures the impact of knowledge updates on graph connectivity and global information propagation, providing the structural basis for detecting suspicious knowledge evolution.

R2: Structural Evasion versus Attack Effectiveness (W1/Q1)

Thank you for this observation. While attackers may reduce structural changes to evade Stage I, successful poisoning requires malicious knowledge to survive extraction and integration, be retrieved, and ultimately influence the generated answer. Therefore, structural evasion does not necessarily translate into successful end-to-end poisoning.

To investigate this trade-off, we compare two poisoning strategies: (1) Regular Poisoning, which injects malicious documents without explicitly restricting structural changes, and (2) Relation Rewiring, which uses black-box queries to infer potentially existing entities and introduces misleading relationships to reduce structural perturbations. We evaluate both strategies using LightRAG with Llama-3.1-8B on Graph-Novel and HotpotQA, with 50 attack attempts per strategy and dataset under the original black-box threat model. We use comparable attack targets and budgets and retain the fixed threshold τ=3×10−4\tau=3\times10^{-4}. Detection measures the proportion of malicious document updates rejected by Stage I, while ASR is calculated over all attack attempts.

DatasetAttackDetection ↑Vanilla ASR ↑Stage I-only ASR ↓SafeG ASR ↓
Graph-NovelRegular76.0046.0020.0012.00
Graph-NovelRelation Rewiring34.0024.0020.0012.00
HotpotQARegular72.0044.0018.0010.00
HotpotQARelation Rewiring32.0026.0022.0012.00

Relation Rewiring reduces the average Stage I detection rate from 74.00% to 33.00%, while its ASR against undefended GraphRAG decreases from 45.00% to 25.00%. These results reveal a trade-off between structural evasion and attack effectiveness. Moreover, full SafeG further reduces Relation Rewiring ASR to 12.00%, demonstrating its effectiveness against structurally adaptive poisoning.

R3: False Rejections of Legitimate Knowledge Updates (W1/Q1)

Thank you for highlighting the importance of preserving legitimate knowledge updates. To evaluate whether structural novelty causes false rejections, we additionally examine high-impact benign updates, defined as the top 20% of genuinely new benign documents ranked by structural impact.

DatasetOverall Benign FPR ↓High-impact Benign FPR ↓
Graph-Novel23.00%32.50%
HotpotQA20.50%27.50%
Average21.75%30.00%

The results show that high-impact benign updates have a higher false-rejection rate than the overall benign population. Nevertheless, Stage I accepts 70.00% of these structurally novel benign updates on average, indicating that substantial structural changes do not necessarily trigger rejection. These findings quantify the trade-off between detecting suspicious updates and preserving legitimate knowledge evolution.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/2)

R4: Cumulative Structural Deviations (W2/Q2)

Thank you for identifying this issue. Since Stage I evaluates each document against the same pre-update KG within a batch, individually accepted documents may collectively introduce substantial structural changes. To investigate this, we conducted a stress test using Split-Payload Poisoning, which distributes malicious knowledge across multiple documents to reduce individual structural deviations.

We evaluated this strategy using LightRAG with Llama-3.1-8B on Graph-Novel and HotpotQA, with 50 attack attempts per dataset and τ=3×10−4\tau=3\times10^{-4}, under the black-box threat model. We measure All Documents Accepted, the proportion of attacks whose documents all pass Stage I, and Cumulative JS > τ\tau, the proportion of accepted batches whose combined structural deviation exceeds the threshold.

DatasetAll Documents AcceptedCumulative JS > τ\tau
Graph-Novel32/50 (64.00%)28/32 (87.50%)
HotpotQA34/50 (68.00%)31/34 (91.18%)
Overall66/100 (66.00%)59/66 (89.39%)

The results show that 89.39% of fully accepted batches produce cumulative structural deviations exceeding τ\tau, confirming that individually accepted updates can collectively induce substantial structural changes. This experiment characterizes the compositional behavior of document-level screening and identifies the need to evaluate its downstream security impact.

R5: Effectiveness Against Cumulative Poisoning (W2/Q2)

To assess the security impact of Split-Payload Poisoning, we compare it with Regular Poisoning under matched attack targets and budgets. We evaluate Vanilla GraphRAG, Stage I alone, and full SafeG, using 50 attacks per strategy and dataset. Batch Pass measures the proportion of attacks in which all documents pass Stage I, while ASR measures successful answer manipulation across all attack attempts.

DatasetAttackBatch Pass ↑Vanilla ASR ↑Stage I-only ASR ↓SafeG ASR ↓
Graph-NovelRegular24.0046.0020.0012.00
Graph-NovelSplit-Payload64.0032.0028.0014.00
HotpotQARegular28.0044.0018.0010.00
HotpotQASplit-Payload68.0028.0022.0016.00

Compared with Regular Poisoning, Split-Payload Poisoning increases Batch Pass and Stage I-only ASR on both datasets, despite lower Vanilla ASR. Full SafeG reduces Split-Payload ASR from 28.00% to 14.00% on Graph-Novel and from 22.00% to 16.00% on HotpotQA, demonstrating the complementary protection of Stage II against cumulative poisoning.

R6: Reliability of Parametric Knowledge in Stage II (Q3)

Thank you for your question. Stage II generates candidate answers from retrieved KG evidence and uses independently generated parametric knowledge as an auxiliary reference for credibility verification. To examine its reliability when the LLM's internal knowledge is insufficient or incorrect, we conducted experiments using LightRAG with Llama-3.1-8B.

(1) We first selected queries where No-GraphRAG answers incorrectly but clean GraphRAG answers correctly, focusing on cases where external knowledge is essential for accurate responses.

Dataset# QueriesClean GraphRAG ACC ↑+ Stage II ACC ↑
Graph-Medical32100.00%84.38%
Graph-Novel36100.00%83.33%
HotpotQA46100.00%84.78%
2WikiMultihopQA43100.00%81.40%
Overall157100.00%83.44%

Stage II preserves 131 of 157 correct answers (83.44%), demonstrating its ability to retain KG-supported answers even when the base LLM fails without retrieval.

(2) We further examined the generated knowledge references on Graph-Medical and Graph-Novel, distinguishing Unknown Prior, where the reference lacks necessary facts, from Incorrect Prior, where it contradicts the ground truth.

DatasetPrior Condition# QueriesStage II CorrectStage II ACC ↑
Graph-MedicalUnknown Prior181688.89%
Graph-MedicalIncorrect Prior141178.57%
Graph-NovelUnknown Prior211780.95%
Graph-NovelIncorrect Prior151386.67%
OverallUnknown Prior393384.62%
OverallIncorrect Prior292482.76%

Stage II preserves 84.62% and 82.76% of correct KG-supported answers under unknown and incorrect parametric knowledge, respectively. These results demonstrate that credibility verification can retain most valid external knowledge even when the generated reference is incomplete or inaccurate, while also quantifying the remaining impact of unreliable priors.

We sincerely thank you again for the careful evaluation and constructive suggestions. We believe these additional experiments and clarifications have strengthened the manuscript.

Thank you for the thorough experiments and candid analysis in R1–R6.

The Relation Rewiring (R2) and Split-Payload (R4–R5) experiments directly validate my intuition regarding structural evasion and cumulative batch deviations in Stage I, while convincingly showing how Stage II compensates for those failure modes in end-to-end ASR. Likewise, the benign FPR results in R3 (21.75% overall, 30.00% on high-impact updates) and the unknown/conflicting prior evaluation in R6 (~83.4% retention, ~16.6% drop) provide an honest and concrete quantification of the system's deployment trade-offs.

Most of my questions are addressed. Please incorporate R1–R6 into the revised paper so readers have a clear view of both the complementarity of the two stages and their practical operating costs. I will maintain my positive rating.

We sincerely thank you for your thoughtful feedback and positive assessment. We greatly appreciate your recognition of our additional experiments and analyses. We will incorporate these improvements into the revised manuscript.

Reviewer_bv3rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SafeG, a two-stage defense against poisoning attacks on GraphRAG. During KG construction, each incoming document is merged into the existing KG, and PageRank-style stationary distributions are computed before and after the merge. The document is dropped if their JS divergence exceeds a threshold. At inference time, the LLM answers the query using nested subgraphs retrieved with top-k = 1..K. It then writes a background passage from its own knowledge and uses it to score the candidates and pick the best one. Experiments cover 4 datasets, 3 GraphRAG frameworks, 3 LLMs and two attacks (KEPo, GRAG-Poison). SafeG reports higher ACC and lower ASR than 11 baselines.

Critical strengths
  1. Defending GraphRAG at ingestion time is underexplored; the existing GraphRAG-specific defense (HoG-GRAG) only acts after the poison is already in the graph. Using structural change rather than text features is a reasonable idea, since attack texts are written to be fluent. It is also cheap, adding about 0.45 s per document.
  2. The evaluation is broad (3 frameworks × 3 LLMs × 4 datasets × 2 attacks, 11 baselines), and the gains are consistent, so the results do not depend on one particular setup.
Critical weaknesses
  1. Novelty is limited. Eq. 3 is standard PageRank, so Stage I is PageRank + JS divergence + a threshold. Stage II is essentially multi-top-k retrieval (close to the Expansion baseline), plus the internal-knowledge idea from AstuteRAG, plus LLM-as-judge. The two stages do not interact, so the "lifecycle" framing is really two filters stacked together. Table 1 suggests neither stage alone beats strong baselines (Stage I alone is below HoG-GRAG), so most of the gain seems to come from the stacking.
  2. Stage I's score measures novelty, not maliciousness. With uniform teleportation and zero mass for new nodes, the divergence mostly grows with how many new entities a document adds. As a result, benign documents carrying new knowledge get flagged (about 20% FPR at about 70% recall in Fig. 11). A fixed absolute τ also does not suit a KG whose size keeps changing. And because documents are scored one at a time, poison split across several documents would not be caught.
  3. Stage II uses the LLM's parametric knowledge as the judge of truth, which goes against the reason for using RAG in the first place. It cannot tell poison from a genuine knowledge update (KEPo poisons are written as "updates"), so in practice it falls back to the model's prior. The subgraphs are also nested, so a poison ranked top-1 appears in every view, and the multi-view design gives no real isolation.
  4. There are no adaptive attacks and no clean (no-attack) utility results. An attacker who only rewires relations between existing entities, or who splits the payload across documents, would likely bypass Stage I. The cost of dropping benign documents is never measured.
Questions for authors
  1. Are the benign incoming documents drawn from the same corpus used to build the clean KG? If so, they add almost no new structure, and Stage I's separation would be inflated. Please report results with genuinely new benign documents, along with no-attack ACC.
  2. How does Stage I hold up against an adaptive attacker, e.g., one that rewires relations among existing entities or splits the payload across several documents that each stay below τ?
  3. How does Stage II behave when the correct answer conflicts with the LLM's prior, e.g., on Graph-Novel, or on queries where No-GraphRAG fails but clean GraphRAG succeeds?
  4. Is τ = 3×10⁻⁴ shared across all frameworks despite their very different graph sizes? And how does Stage I combined with an existing defense (e.g., HoG-GRAG) compare with full SafeG?
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/3)

Thank you very much for your careful and constructive review. We greatly appreciate your recognition that "defending GraphRAG at ingestion time is underexplored," that "using structural change rather than text features is a reasonable idea," and that "the evaluation is broad" with consistent improvements across different settings. We also greatly appreciate your insightful suggestions for further strengthening our work. We will address your comments and questions individually in the following ten parts:

R1: Methodological Novelty (W1)

Thank you for your comment. Existing GraphRAG defenses primarily focus on reasoning-time protection, leaving malicious knowledge updates insufficiently addressed during KG evolution. Moreover, ingestion-time detection alone cannot fully prevent poisoned knowledge from affecting downstream reasoning. SafeG addresses these two interconnected challenges through a lifecycle-oriented defense that jointly protects knowledge evolution and answer generation.

Specifically, Stage I identifies suspicious knowledge updates by measuring their impact on global information propagation, while Stage II mitigates residual poisoning by generating and verifying candidate answers across multiple evidence scopes. These mechanisms address distinct failure points and complement each other, rather than repeatedly filtering the same information. The ablation results in Table 1 demonstrate the contribution of both stages, while the additional comparison with Stage I + HoG-GRAG in R10 further validates the effectiveness of SafeG's integrated design.

R2: Interpretation of Structural Deviation (W2)

Thank you for your comment. Stage I evaluates the impact of incoming knowledge on global information propagation within the KG, capturing structural changes through shifts in the stationary distribution. While both benign and poisoned updates may introduce new entities and relations, poisoned updates often establish abnormal associations that produce more pronounced changes in global information propagation.

This distinction is supported by the experimental results in Figure 10 (page 20), where Stage I achieves AUROC values of 0.847, 0.769, 0.814, and 0.797 across four datasets. These results demonstrate the effectiveness of structural deviation as a discriminative signal for identifying suspicious knowledge updates.

R3: Threshold Generalizability (W2/Q4)

Thank you for your question. To evaluate the applicability of the detection threshold across different KG sizes, we use a unified threshold of τ=3×10−4\tau=3\times10^{-4}, selected through validation, across all frameworks and datasets.

As shown in Table 2, the evaluated KGs vary substantially in size, ranging from 1,487 to 9,324 entities and 981 to 68,622 edges. The consistent experimental performance across these settings supports the practical applicability of the selected threshold, while Figure 6(b) further examines its sensitivity.

R4: Reliability of Parametric Knowledge (W3/Q3)

Thank you for your question. Stage II uses an independently generated knowledge reference as an auxiliary signal for credibility assessment, while candidate answers are generated from retrieved KG evidence. To examine whether this mechanism can preserve correct answers when the LLM's parametric knowledge is insufficient or incorrect, we conducted additional experiments using LightRAG with Llama-3.1-8B. Specifically, we selected queries for which No-GraphRAG answered incorrectly but clean GraphRAG answered correctly, and evaluated whether Stage II retained the correct KG-supported answers.

Dataset# QueriesClean GraphRAG ACC ↑+ Stage II ACC ↑
Graph-Medical32100.00%84.38%
Graph-Novel36100.00%83.33%
HotpotQA46100.00%84.78%
2WikiMultihopQA43100.00%81.40%
Overall157100.00%83.44%

The 100.00% clean GraphRAG accuracy follows from the query-selection criterion. Stage II retains 131 of 157 correct answers (83.44%), demonstrating its ability to preserve KG-supported answers even when the base model fails without retrieval. We further examined 68 queries from Graph-Medical and Graph-Novel. Stage II retained 33/39 correct answers (84.62%) when the knowledge reference lacked the required facts and 24/29 (82.76%) when it contradicted the ground truth. These results demonstrate the effectiveness of Stage II under incomplete or incorrect parametric knowledge while quantifying its remaining limitations.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/3)

R5: Effectiveness of Nested Retrieval Views (W3)

Thank you for this observation. Stage II employs nested retrieval views to generate candidate answers from different evidence scopes, rather than relying on independent or disjoint subgraphs. By varying the retrieval budget from top-1 to top-K, the mechanism incorporates progressively broader KG evidence, which may introduce additional supporting or conflicting information. The resulting candidates are evaluated through credibility verification, allowing SafeG to compare answers under different knowledge contexts. The effectiveness of this design lies in evidence-dependent candidate comparison and verification, complementing Stage I's ingestion-time screening against residual poisoning.

R6: Robustness Against Adaptive Attacks (W2/W4/Q2)

Thank you for your suggestion. Adaptive attacks may reduce structural deviations to evade detection, but such constraints may also limit their effectiveness in manipulating final answers. To investigate this trade-off, we conducted additional experiments with two adaptive strategies: (1) Relation Rewiring, which uses black-box queries to infer potentially existing entities and inject misleading relationships; and (2) Split-Payload Poisoning, which distributes malicious knowledge across multiple documents to reduce per-document structural deviations.

We evaluated both attacks on Graph-Novel and HotpotQA using LightRAG with Llama-3.1-8B, with 50 attack attempts per strategy and dataset under the original black-box threat model.

DatasetAttackVanilla ACC ↑Vanilla ASR ↓SafeG ACC ↑SafeG ASR ↓
Graph-NovelRelation Rewiring54.0024.0062.0012.00
Graph-NovelSplit-Payload48.0032.0058.0014.00
HotpotQARelation Rewiring56.0026.0062.0012.00
HotpotQASplit-Payload50.0028.0060.0016.00

SafeG reduces average ASR from 25.00% to 12.00% under Relation Rewiring and from 30.00% to 15.00% under Split-Payload Poisoning, while improving ACC by 7.00 and 10.00 percentage points, respectively. These results demonstrate SafeG's effectiveness against both adaptive attacks and highlight that structural evasion does not necessarily translate into successful answer manipulation. We have incorporated the additional experimental results and analysis into Appendix E.

R7: Clean Utility under Benign Knowledge Updates (W4)

Thank you for your suggestion. To quantify the utility cost of incorrectly rejecting benign documents, we conducted an additional experiment under attack-free KG evolution, focusing on how Stage I affects downstream answer accuracy when all incoming documents are benign.

We evaluated Graph-Novel and HotpotQA using LightRAG with Llama-3.1-8B. For each dataset, we considered 200 benign incoming documents and 200 evaluation queries, without introducing any poisoned documents. Starting from the same initial KG, we compared Vanilla (Clean), which incorporates all benign documents, with Stage I (Clean), which filters incoming documents using the original threshold τ=3×10−4\tau=3\times10^{-4}. We report the benign false positive rate (FPR) and answer accuracy (ACC) to quantify the impact of benign document rejection.

DatasetBenign FPR ↓Vanilla ACC ↑Stage I ACC ↑ACC Drop ↓
Graph-Novel23.0057.5053.004.50
HotpotQA20.5053.5050.003.50
Average21.7555.5051.504.00

The results show that Stage I rejects 21.75% of benign documents on average, resulting in an ACC reduction of only 4.00 percentage points. Despite these false positives, most downstream queries remain correctly answered, indicating that document-level rejection does not translate proportionally into answer-level utility loss. These results directly quantify the clean-utility trade-off introduced by Stage I, demonstrating that its impact on overall answer accuracy remains relatively limited.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3/3)

R8: Genuinely New Benign Documents (Q1)

Thank you for pointing out this ambiguity. In Section 5.1, the "original corpus" refers to the original benchmark corpus, rather than the documents used to construct the initial clean KG. Benign incoming documents were excluded from initial KG construction and subsequently introduced during KG evolution. Therefore, the evaluated benign documents represent new knowledge inputs rather than documents already incorporated into the KG. We have clarified the experimental setup in the revised manuscript:

In the poisoning experiments, benign documents are sampled from the original benchmark corpus but are excluded from the initial KG construction. They are subsequently introduced as new documents during KG evolution, mixed with poisoned documents at a 1:1 ratio with balanced lengths.

R9: Clean Performance (W4/Q1)

Thank you for your suggestion. To provide a clean-performance reference, we additionally evaluated the ACC of three GraphRAG frameworks across four datasets without poisoning. In this setting, all incoming benign documents are directly integrated into the KG without Stage I screening.

FrameworkGraph-MedicalGraph-NovelHotpotQA2WikiMultihopQAAverage
LightRAG66.1763.5063.5063.6764.21
FastGraphRAG66.3358.1761.5061.8361.96
HippoRAG269.5060.5064.0064.3364.58
Average67.3360.7263.0063.2863.58

These results establish a clean-performance reference across different GraphRAG frameworks.

R10: Comparison with HoG-GRAG (W1/Q4)

Thank you for your suggestion. To assess the additional contribution of SafeG's Stage II beyond combining Stage I with an existing reasoning-stage defense, we conducted additional experiments comparing Stage I + HoG-GRAG with full SafeG under KEPo attacks using LightRAG with Llama-3.1-8B across four datasets.

DatasetStage I + HoG-GRAG ACC ↑ASR ↓Full SafeG ACC ↑ASR ↓
Graph-Medical59.3317.6761.6714.67
Graph-Novel55.6718.6757.1715.00
HotpotQA54.5021.0057.6717.67
2WikiMultihopQA56.6719.3359.1716.00
Average56.5419.1758.9215.84

Full SafeG consistently outperforms Stage I + HoG-GRAG, improving average ACC by 2.38 percentage points and reducing ASR by 3.33 percentage points. This improvement may stem from HoG-GRAG's reliance on hop-wise inconsistency detection, which can overlook coherent but misleading evidence, whereas SafeG's Stage II evaluates candidate answers across multiple evidence scopes through credibility verification.

We sincerely thank you again for the careful evaluation and constructive suggestions. We believe these additional experiments and clarifications have strengthened the manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear AC and Reviewers,

We sincerely thank all reviewers for their constructive feedback and active engagement. All three reviewers (bv3r, kUAo, nnvX) have responded to our rebuttal and now recommend acceptance. Reviewers bv3r and nnvX raised their scores after the rebuttal, and Reviewer kUAo maintained a positive rating.

We would also like to briefly highlight the significance of our work. GraphRAG is vulnerable to poisoning attacks that contaminate its knowledge graph, yet existing defenses mainly target conventional RAG or only repair already-poisoned KGs. SafeG fills this gap with a novel lifecycle-oriented defense spanning both KG evolution and inference, which the reviewers found well motivated, lightweight, and consistently effective across extensive settings.

In the rebuttal, we addressed each concern raised by the reviewers through additional experiments and clarifications, which the reviewers acknowledged in their responses. We will add all of these results to the revised manuscript and, as Reviewer nnvX suggested, strengthen the theoretical analysis.

Thank you again for your valuable time and support.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.