acceptodds
Under review as a conference paper at ICLR 2027

SILK: Semantic Network-Driven Iterative Query Rewriting for RAG Knowledge Extraction

Abstract

Retrieval-Augmented Generation (RAG) systems ground large language model responses in external knowledge bases, making these costly proprietary databases attractive targets for knowledge extraction attacks. Existing explicit extraction methods achieve high stealing rates but are vulnerable to defenses, while implicit methods maintain stealth at the cost of substantially lower stealing rates. To resolve this dilemma, we propose SILK, which consists of two modules: anchor selection and query generation. For anchor selection, SILK constructs a semantic network of topic anchors and decomposes RAG responses into atomic claims, enabling fine-grained knowledge gain measurement; it further leverages the local continuity of knowledge density in the embedding space to propagate observed scores from explored nodes to estimate the potential of unexplored anchors. For query generation, SILK formulates the task as a constrained optimization problem, where a shadow intent detector serves as a hard constraint pruning infeasible queries and a judge-guided constrained beam search iteratively rewrites queries to maximize leakage within the feasible region. Experiments on four datasets demonstrate that SILK consistently outperforms existing six baselines under various defense and RAG configurations, and improves the leakage rate by 17% on average compared to the state-of-the-art implicit method. These results point to a pressing need for stronger protection of RAG proprietary knowledge.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
RAG Knowledge Extraction Attack, RAG Privacy Protection

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_yb3rOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SILK, a framework for implicit knowledge extraction from RAG systems through a black-box query interface. SILK combines two components: semantic-network-based anchor selection and constrained query rewriting. The first component uses atomic claims to estimate knowledge gain, propagates observed scores to unexplored neighboring anchors, and balances topic continuity with exploration. The second component uses a shadow intent detector and an LLM judge to guide constrained beam search over candidate queries.

The evaluation compares SILK with six baselines across four datasets, multiple defense settings, and alternative model configurations. The paper additionally includes component ablations, evaluator validation, cost analysis, and an evaluation of surrogate RAG systems constructed from the extracted knowledge.

Critical strengths
  1. A coherent framework for tracking and expanding knowledge acquisition in the implicit setting. Measuring extraction progress is difficult when responses do not reproduce retrieved passages verbatim. SILK addresses this through atomic-claim-based knowledge tracking and uses a semantic network to guide exploration. The combination is well motivated, and the ablations in Section 5.6 support the usefulness of its components. The results also help distinguish the contribution of anchor selection to retrieval coverage from that of query rewriting to leakage.

  2. Consistent empirical improvements over the relevant implicit baseline. SILK outperforms IKEA on EE, LR, and SR across all four datasets under both input and output defenses in Table 1. For example, under input defense on FIQA, SR increases from 0.43 to 0.71. The additional evaluations with different attack models and RAG configurations provide further evidence that the improvements are not restricted to one configuration. Together with the surrogate-RAG evaluation and cost measurements, these results constitute a useful empirical contribution.

Critical weaknesses
  1. The interpretation of semantic similarity as knowledge leakage needs qualification. Section 5.1 classifies a retrieved chunk as leaked when either ROUGE-L recall exceeds 0.5 or semantic similarity exceeds 0.65. The semantic criterion is useful for capturing paraphrases, but it may also count topical similarity without establishing recovery of distinctive facts. Appendix B provides helpful human validation, although the leakage assessment covers only 30 pairs. Moreover, the surrogate-RAG evaluation measures response similarity rather than factual correctness or the recovery of information unavailable without access to the target knowledge base. These limitations do not negate the reported comparative gains, but they make the stronger claims about proprietary knowledge recovery and near-equivalent service reconstruction less conclusive.

  2. The defense evaluation and uncertainty reporting need clearer specification. Section 5.4 does not sufficiently describe the attack composition, train/test separation, sequence construction, and calibration protocol for the sequence detector. The model configurations and relationship between the shadow and target intent detectors also require clarification. In addition, the results do not report variability despite stochastic anchor sampling and LLM generation. These omissions limit reproducibility and make it difficult to determine how broadly the low blocking rates should generalize beyond the evaluated implementations.

Questions for authors
  1. What performance is obtained when access to the target knowledge base is removed, both for the leakage metric and for the surrogate-system evaluation? Such a control would help distinguish knowledge-base-specific recovery from information already available in the underlying model. Can the authors also clarify how the semantic-similarity threshold was selected and how the 30 human-evaluated leakage samples were distributed across datasets and methods?

  2. Please clarify the configurations of the shadow and target intent detectors, the three attacks used to train the sequence detector, the train/test separation, and the construction of the evaluated sequences. How was the 5% false-positive operating point calibrated? How many independent runs were conducted, and how variable were the principal results?

  3. Section 5.1 defines SR = EE × LR, but the PoR result on HealthCareMagic without defense in Table 1 reports EE = 0.72, LR = 0.79, and SR = 0.64, whereas the product is approximately 0.57. Does this arise from separate averaging, another aggregation procedure, or a reporting error? Please also reconcile the discrepancies between Appendix A and Tables 6–7.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

Response to Weakness 1 / Question 1: Leakage validity, surrogate controls, and human evaluation

Regarding the source of the extracted knowledge

A primary goal of RAG is to provide domain-specific knowledge while reducing hallucinations [1, 2, 3]. In line with this goal, our RAG setting follows a standard enterprise configuration used in prior work [4, 5, 6] (Section 5.1; see also Prompt 1, Appendix I\color{blue}{\text{Prompt 1, Appendix I}}): the model must answer strictly based on retrieved content, and respond “I don’t know” when the retrieved evidence is insufficient. Thus, in our experiments, the LLM is intended to reason over retrieved evidence rather than answer from parametric knowledge.

To verify this, we conduct two retrieval-control experiments using the actual query sets from each attack: (1) retrieval disabled; and (2) retrieved content unrelated to the question. We have added the results of these retrieval-control experiments in Tables 9–10, Appendix B.1\color{blue}{\text{Tables 9–10, Appendix B.1}}. We test DeepSeek-V3 and GLM-4.7. Under both conditions, both models consistently answer “I don’t know”, with a leakage rate of 0, rather than relying on pretrained knowledge. This shows that the extracted content comes from the knowledge base, not the model’s parametric knowledge. Therefore, Table 1 measures knowledge base leakage, supporting our claim that SILK is highly effective at extracting RAG knowledge bases.

Regarding the surrogate-system evaluation

To isolate the effect of the knowledge base on RAG performance, we keep all other components fixed when comparing the surrogate and original RAG systems. They use the same backend, retrieval configuration, and answering prompt; only the knowledge base changes. We additionally disable retrieval for the surrogate while retaining the evaluation queries. As shown in Table 17, Appendix B.4\color{blue}{\text{Table 17, Appendix B.4}}, the surrogate then consistently answers “I don’t know,” with ROUGE-L and semantic similarity both below 0.01 across all four datasets. With retrieval enabled, the SILK surrogate exceeds the IKEA surrogate on both measures. This control supports the contribution of the extracted KB to surrogate performance.

Regarding the semantic threshold

We add a semantic branch because implicit extraction can preserve information while changing its wording. We now evaluate five cutoffs on the same 30 human-annotated pairs, keeping the ROUGE-L recall threshold fixed at 0.5. Among the tested cutoffs, 0.65 gives the highest observed accuracy, F1, and κ (Table 18, Appendix B.5\color{blue}{\text{Table 18, Appendix B.5}}).

Cosine cutoffAccuracyF1κ
0.750.7670.7410.553
0.700.8000.8130.600
0.650.8330.8570.658
0.600.7670.8210.502
0.550.7330.8100.415

Regarding annotation coverage

Table 19, Appendix B.5\color{blue}{\text{Table 19, Appendix B.5}} provides the sample counts for every method–dataset combination. Under No Defense, we randomly sample 30 retrieved chunks from the pooled retrievals of all seven methods across all four datasets, pairing each chunk with its corresponding response. The dataset totals are 7/8/6/9 for HealthCareMagic/FIQA/SciFact/NFCorpus. The method totals are 6/3/3/2/4/6/6 for TGTB/PIDE/DGEA/RAG-Thief/PoR/IKEA/SILK, respectively.


AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

Response to Weakness 2 / Question 2: Detector configuration, calibration, and variability

Regarding the target and shadow intent detectors

Under our threat model, the target detector’s prompt is unknown to the attacker. We therefore use different LLMs and prompts for the target and shadow intent detectors: the target detector uses DeepSeek-V3, and the shadow detector uses GLM-4.7. Their prompts are provided under RAG Intent Detection and Shadow Intent Detection in Appendix I. These configurations are now explicit in Appendix F.4\color{blue}{\text{Appendix F.4}}.

Regarding sequence construction and train/test separation

Our sequence-detector evaluation follows the sequence-detection setup in IKEA [7], with benign dialogue generation using DialogueForge [8]. The 2,200 benign dialogues use OASST1 and Chatbot Arena seeds, with one seed user query followed by five synthesized user queries. Each dialogue forms one six-query sequence. We select PoR, IKEA, and SILK as the three training attack sources to cover a strong explicit extraction attack and the two stealthy implicit attacks. We collect two 200-round trajectories per attack and dataset: 24 trajectories in total. Six-query windows with a stride of three produce 65 sequences per trajectory. We use a chronological split within each trajectory, retaining 51 earlier training sequences and 13 later test sequences, and discarding one boundary window. The additional attack test set contains 100 sequences each from TGTB, PIDE, DGEA, and RAG-Thief.

The data composition is provided in Table 31, Appendix F.5\color{blue}{\text{Table 31, Appendix F.5}}:

SourceTrainingCalibrationTest
PoR / IKEA / SILK1,2240312
TGTB / PIDE / DGEA / RAG-Thief00400
Benign dialogues1,600200400
Total sequences2,8242001,112

Regarding the 5% false-positive operating point

We calibrate the sequence detector on the 200 held-out benign dialogues, selecting the smallest evaluated threshold with empirical query-level FPR below 5%, then freezing it for testing. A query is blocked if any evaluated window containing it is flagged. At the selected threshold, nine calibration dialogues are flagged. Each is one six-query window, corresponding to 54/1,200 queries, or 4.5% FPR. The equations and aggregation procedure are given in Appendix F.5\color{blue}{\text{Appendix F.5}}.

Regarding independent runs and variability

Each reported extraction result comes from one complete attack trajectory (N=200N=200 rounds) per method, dataset, and setting. The paper evaluates SILK under 14 distinct settings per dataset: the three principal defense settings (Table 1) and eleven additional settings covering alternative attack LLMs, backend LLMs, retrievers, and rerankers (Tables 20–23). We summarize the 14 independently executed runs per dataset in Table 25, Appendix D.4\color{blue}{\text{Table 25, Appendix D.4}}: the standard deviations of EE, LR, and SR are all below 0.05 (maximum: 0.043), indicating SILK’s robustness.

Moreover, SILK exceeds IKEA in all 48 paired configuration–defense–dataset comparisons, with mean ΔSR +19.9%. We consider that these results account not only for the randomness of stochastic anchor sampling and LLM generation, but also for different RAG configurations and attack models; together they indicate that SILK is stable with respect to both randomness and configuration.

SILK’s EE, LR, and SR across the 14 settings per dataset are summarized below as mean ± sample SD (Table 25, Appendix D.4\color{blue}{\text{Table 25, Appendix D.4}}).

DatasetEE (mean ± SD)LR (mean ± SD)SR (mean ± SD)
HealthCareMagic0.813 ± 0.0280.694 ± 0.0290.564 ± 0.036
FIQA0.854 ± 0.0330.806 ± 0.0300.689 ± 0.043
SciFact0.644 ± 0.0260.601 ± 0.0190.386 ± 0.023
NFCorpus0.577 ± 0.0350.685 ± 0.0250.395 ± 0.033

The paired comparison with IKEA is summarized below (Table 26, Appendix D.4\color{blue}{\text{Table 26, Appendix D.4}}).

MetricPairsMean difference (SILK − IKEA)SDSILK higher
SR48+0.1990.05948/48
LR48+0.1410.06048/48

Response to Question 3: SR consistency and Appendix A discrepancies

Comparison with the original records confirmed that the PoR SR discrepancy was a reporting error. We corrected the entry and carefully checked the other reported results for the same issue. We have also corrected the discrepancies between Appendix A and Tables 6–7 and checked the corresponding descriptions and tables throughout the manuscript for consistency.


AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

References

[1] Lewis, Patrick, et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS (2020). Paper.

[2] Gao, Yunfan, et al. “Retrieval-Augmented Generation for Large Language Models: A Survey.” arXiv:2312.10997 (2023). Paper.

[3] Shuster, Kurt, et al. “Retrieval Augmentation Reduces Hallucination in Conversation.” Findings of EMNLP (2021). Paper.

[4] Leung, Kin Kwan, et al. “Classifying and Addressing the Diversity of Errors in Retrieval-Augmented Generation Systems.” EACL (2026). Paper.

[5] Zhang, Maosen, et al. “LeakDojo: Decoding the Leakage Threats of RAG Systems.” Findings of ACL (2026). Paper.

[6] Di Maio, Christian, et al. “Pirates of the RAG: Adaptively Attacking LLMs to Leak Knowledge Bases.” arXiv:2412.18295 (2024). Paper.

[7] Wang, Yuhao, et al. “Silent Leaks: Implicit Knowledge Extraction Attack on RAG Systems through Benign Queries.” ICLR (2026). Paper.

[8] Zhu, Ruizhe, et al. “DialogueForge: LLM Simulation of Human-Chatbot Dialogue.” arXiv:2507.15752 (2025). Paper.

AuthorsOpenReview ↗

Looking forward to further feedback

Dear Reviewer yb3r,

As the rebuttal deadline approaches, we would like to kindly follow up on our response and welcome any further feedback. If our clarifications address your concerns, we would greatly appreciate your reconsideration of the rating.

Thank you very much for your support!

Best,

The Authors

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

This paper studies implicit knowledge extraction from black-box retrieval-augmented generation systems. SILK organizes query anchors using a semantic network, guides exploration through the novelty and repetition of extracted atomic claims, and rewrites queries under semantic-anchor constraints and a shadow detector. The objective is to increase knowledge coverage and extraction success within a limited query budget while avoiding detection. The evaluation includes four datasets, multiple RAG configurations, detection-based defenses, downstream performance of a surrogate RAG system, component ablations, and operational costs.

Critical strengths

The threat model is practically relevant. A sequence of individually innocuous queries may cumulatively extract substantial information, even when single-turn intent detection performs reasonably well. Combining coverage-oriented exploration with query-level concealment is therefore a meaningful direction.

The empirical evaluation is broader than the main results alone suggest. In addition to input- and output-side defenses, the appendices examine sequence-level and semantic-drift detection, alternative model and retrieval configurations, human validation of evaluation metrics, and running costs. These experiments provide useful context for the proposed attack and should be credited when assessing its completeness (Section 5; Appendices B, D, and F).

Critical weaknesses

W1. The leakage criterion does not sufficiently establish that the extracted information originates from the target knowledge base.

Leakage is identified using ROUGE-L recall above 0.5 or semantic similarity above 0.65. With publicly available domain corpora, answers based on parametric knowledge, common domain facts, or semantically similar material may satisfy these thresholds without demonstrating source-specific extraction. Likewise, atomic-claim novelty measures what the attacker has newly observed, not necessarily what originated from the target corpus. The surrogate-RAG results support the usefulness of the collected material but do not independently resolve this attribution issue (Section 5.1–5.3, p. 7; Appendix B, p. 14). A focused control would substantially strengthen the evidence: evaluate a fixed subset of queries with retrieval disabled or the corresponding source documents removed, or use controlled facts that the model is unlikely to know beforehand. Independent claim-to-source verification would further establish whether SILK’s gains reflect access to the target knowledge base rather than related content generation.

W2. Claim repetition should be interpreted as a coverage heuristic rather than an identified estimate of corpus coverage.

Repeated-claim frequency is a plausible signal of exploration saturation. Under adaptive querying and nonuniform retrieval, however, it is not equivalent to the fraction of the underlying knowledge base that has been recovered. The ablations support the usefulness of the combined procedure but do not establish when this coverage proxy is reliable or misleading (Section 4; Section 5.6). The paper would benefit from validating the proxy against observable unique-document or ground-truth fact coverage in the experimental setting, particularly in cases where queries repeatedly reach a popular subset of the corpus.

Questions for authors
  1. Can you provide the TP/FP/FN/TN counts, annotator comparisons, aggregation procedures, and κ calculations underlying Table 8? Please also clarify how EE, LR, and SR are aggregated in Table 1.
  2. For a fixed query subset, how often does the current leakage criterion remain positive when retrieval is disabled or the relevant documents are removed? Does SILK retain its advantage after source attribution is verified?
  3. How closely does claim repetition track actual unique-document or fact coverage? Are there cases where repetition becomes high despite substantial unrecovered content?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/2]

Response to Question 1: Validation details for Table 8 and metric aggregation in Table 1

We have added Tables 12–16, Appendix B.3\color{blue}{\text{Tables 12–16, Appendix B.3}}, providing the TP/FP/FN/TN counts, annotator comparisons and aggregation procedures, and detailed Cohen’s κ calculations underlying Table 8, including pairwise rating tables and sample-level annotations. We also clarify the cumulative chunk-level aggregation of EE, LR, and SR in Table 1, including deduplication and rounding.


Response to Weakness 1 / Question 2: Does the extracted knowledge come from the KB or the model’s pretrained knowledge?

A primary goal of RAG is to provide domain-specific knowledge while reducing hallucinations [1, 2, 3]. In line with this goal, our RAG setting follows a standard enterprise configuration used in prior work [4, 5, 6] (Section 5.1; see also Prompt 1, Appendix I\color{blue}{\text{Prompt 1, Appendix I}}): the model must answer strictly based on retrieved content, and respond “I don’t know” when the retrieved evidence is insufficient. Thus, in our experiments, the LLM is intended to reason over retrieved evidence rather than answer from parametric knowledge.

To verify this, we conduct two retrieval-control experiments using the actual query sets from each attack: (1) retrieval disabled; and (2) retrieved content unrelated to the question. We have added the results of these retrieval-control experiments in Tables 9–10, Appendix B.1\color{blue}{\text{Tables 9–10, Appendix B.1}}. We test DeepSeek-V3 and GLM-4.7. Under both conditions, both models consistently answer “I don’t know”, with a leakage rate of 0, rather than relying on pretrained knowledge. This shows that the extracted content comes from the knowledge base, not the model’s parametric knowledge. Therefore, Table 1 measures knowledge base leakage, supporting our claim that SILK is highly effective at extracting RAG knowledge bases.


Response to Weakness 2 / Question 3

Regarding claim repetition as a proxy for document repetition

Unlike explicit attackers, implicit attackers cannot directly observe the retrieved documents. SILK therefore uses claim repetition to estimate document repetition. We have added a validation analysis in Table 11, Appendix B.2\color{blue}{\text{Table 11, Appendix B.2}}, reporting the Spearman correlations and descriptive partial rank correlations controlling for query round. The results show substantial positive correlations, including after controlling for query round, supporting the use of claim repetition as an observable proxy in the implicit setting.

Table 11. Correlations between claim repetition and retrieved-document redundancy (Appendix B.2).\color{blue}{\text{Table 11. Correlations between claim repetition and retrieved-document redundancy (Appendix B.2).}} Partial rank correlations control for query round and are descriptive.

Correlation measureHealthCareMagicFIQASciFactNFCorpus
Spearman correlation (ρ)0.6730.6440.5190.604
Descriptive partial rank correlation (controlling for query round)0.6510.5990.5140.558

Regarding the purpose of the coverage heuristic

We acknowledge that coverage is a heuristic estimate, rather than a formal model of coverage. Under nonuniform retrieval, repetition can become high even while some content within the topic remains unrecovered. However, guaranteeing exhaustive recovery is outside the attacker's objective in our threat model. Attackers typically have a limited query budget, while knowledge bases can be large—for example, HealthCareMagic contains approximately 100,000 records. Their objective is therefore to maximize extraction efficiency within the available budget, rather than exhaust the entire knowledge base. The coverage heuristic is designed to serve this efficiency objective, without guaranteeing that no unrecovered content remains.

Moreover, coverage is only one of the signals guiding SILK's exploration. Anchor selection also considers information increment, the proximity sweet spot, and novelty. Therefore, a high estimated coverage does not rule out selecting an anchor when the other signals indicate substantial exploration value.


AuthorsOpenReview ↗

Rebuttal by Authors [2/2]

References

[1] Lewis, Patrick, et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems 33, pp. 9459–9474 (2020). Paper.

[2] Gao, Yunfan, et al. "Retrieval-Augmented Generation for Large Language Models: A Survey." arXiv preprint arXiv:2312.10997 (2023). Paper.

[3] Shuster, Kurt, et al. "Retrieval Augmentation Reduces Hallucination in Conversation." Findings of the Association for Computational Linguistics: EMNLP 2021, pp. 3784–3803 (2021). Paper.

[4] Leung, Kin Kwan, et al. "Classifying and Addressing the Diversity of Errors in Retrieval-Augmented Generation Systems." Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 3185–3207 (2026). Paper.

[5] Zhang, Maosen, et al. "LeakDojo: Decoding the Leakage Threats of RAG Systems." Findings of the Association for Computational Linguistics: ACL 2026, pp. 5790–5811 (2026). Paper.

[6] Di Maio, Christian, et al. "Pirates of the RAG: Adaptively Attacking LLMs to Leak Knowledge Bases." arXiv preprint arXiv:2412.18295 (2024). Paper.

Thanks for the clarification. I have raised my rating accordingly.

We sincerely appreciate your comments and for raising your rating!

Reviewer_Qnz9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents extraction attacks against the knowledge base of RAG systems. The proposed framework is designed to address robustness against defense methods and high extraction rates. It has two components. In semantic-network anchor selection, SIK identifies promising knowledge topics using previously extracted atomic claims. In constrained query rewriting it uses beam search to generate queries that maximize estimated information leakage. The paper shows improvement over existing methods even with defenses.

Critical strengths
  1. Informative ablation study showing how different components may contribute to the improved coverage.
  2. SIK archives better extraction when defense methods are applied, compared to existing baselines.
  3. Experimental breath is strong, covering 6 baselines, four datasets and two defenses
Critical weaknesses
  1. Lack of control: The paper needs a control experiment to attribute similarity between RAG responses and retrieved chunks to the proprietary knowledge through knowledge retrieval. For the same query, the model could generate similar information from its pretrained knowledge.

  2. Because SIK does not provide consistent improvement over existing baselines without defenses, its contribution seems to be stealthiness or effectiveness even when defenses are in place. However, this makes the paper’s contribution dependent on selected defense methods, and I am not entirely convinced whether the findings generalize beyond these selected methods.

Questions for authors
  1. Does SIK actually extract knowledge bases more effectively than existing baselines or is it simply better at eliciting general knowledge already known to the LLM through carefully crafted queries? To show this, the authors must measure the retrieval-dependent improvement for both SIK and baseline defenses (such as IKEA) and show the differential is higher for SIK than others.
3 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/2]

Response to Weakness 1 / Question 1: Does the extracted knowledge come from the KB or the model’s pretrained knowledge?

A primary goal of RAG is to provide domain-specific knowledge while reducing hallucinations [1, 2, 3]. In line with this goal, our RAG setting follows a standard enterprise configuration used in prior work [4, 5, 6] (Section 5.1; see also Prompt 1, Appendix I\color{blue}{\text{Prompt 1, Appendix I}}): the model must answer strictly based on retrieved content, and respond “I don’t know” when the retrieved evidence is insufficient. Thus, in our experiments, the LLM is intended to reason over retrieved evidence rather than answer from parametric knowledge.

To verify this, we conduct two retrieval-control experiments using the actual query sets from each attack: (1) retrieval disabled; and (2) retrieved content unrelated to the question. We have added the results of these retrieval-control experiments in Tables 9–10, Appendix B.1\color{blue}{\text{Tables 9–10, Appendix B.1}}. We test DeepSeek-V3 and GLM-4.7. Under both conditions, both models consistently answer “I don’t know”, with a leakage rate of 0, rather than relying on pretrained knowledge. This shows that the extracted content comes from the knowledge base, not the model’s parametric knowledge. Therefore, Table 1 measures knowledge base leakage, supporting our claim that SILK is highly effective at extracting RAG knowledge bases.


Response to Weakness 2

Regarding SILK's improvement over existing baselines without defenses

Explicit and implicit extraction differ substantially in difficulty. Explicit methods directly observe retrieved chunks and can therefore model knowledge gain from the full context. Implicit methods only observe the model’s answers, making knowledge gain much harder to infer. Prior implicit methods such as IKEA do not explicitly model this gain, which explains their weaker performance in undefended settings. SILK addresses this by introducing atomic claims to model knowledge gain and guide extraction. As a result, in the much harder implicit setting, SILK achieves performance comparable to, or better than, explicit extraction methods—something prior implicit methods could not achieve.

Meanwhile, although explicit attacks such as PoR are effective without defenses, they are also highly fragile: simple input-side intent detection or rule-based output filtering can block 100% of them. Hence, under common practical defenses, their real-world threat is limited. SILK, by contrast, remains both effective and stealthy, making it the most practical and effective attack currently available.

Regarding SILK's generalizability

Our evaluation already includes a diverse set of defenses, from simple rule-based filters to strong adaptive defenses, covering input-side vs. output-side and single-turn vs. multi-turn settings. These are representative of current RAG defenses and include defenses used in prior work such as IKEA [7] and LeakDojo [5]. Therefore, they were not chosen to favor SILK. The results support SILK’s strong generalization across defense types.

More importantly, SILK is not tailored to any specific defense. It exploits a fundamental constraint of practical defenses: to preserve usability, they cannot tolerate high false positive rates. By mimicking legitimate user behavior, SILK overlaps with normal interactions and becomes difficult to detect without harming benign users. As shown in the newly added Table 24, Appendix F.3\color{blue}{\text{Table 24, Appendix F.3}}, SILK’s queries resemble ordinary information-seeking requests, making it difficult to identify its attack intent even for humans.

AuthorsOpenReview ↗

Rebuttal by Authors [2/2]

References

[1] Lewis, Patrick, et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems 33, pp. 9459–9474 (2020). Paper.

[2] Gao, Yunfan, et al. "Retrieval-Augmented Generation for Large Language Models: A Survey." arXiv preprint arXiv:2312.10997 (2023). Paper.

[3] Shuster, Kurt, et al. "Retrieval Augmentation Reduces Hallucination in Conversation." Findings of the Association for Computational Linguistics: EMNLP 2021, pp. 3784–3803 (2021). Paper.

[4] Leung, Kin Kwan, et al. "Classifying and Addressing the Diversity of Errors in Retrieval-Augmented Generation Systems." Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 3185–3207 (2026). Paper.

[5] Zhang, Maosen, et al. "LeakDojo: Decoding the Leakage Threats of RAG Systems." Findings of the Association for Computational Linguistics: ACL 2026, pp. 5790–5811 (2026). Paper.

[6] Di Maio, Christian, et al. "Pirates of the RAG: Adaptively Attacking LLMs to Leak Knowledge Bases." arXiv preprint arXiv:2412.18295 (2024). Paper.

[7] Wang, Yuhao, et al. "Silent Leaks: Implicit Knowledge Extraction Attack on RAG Systems through Benign Queries." arXiv preprint arXiv:2505.15420 (2025). Paper.

AuthorsOpenReview ↗

Looking forward to further feedback

Dear Reviewer Qnz9,

As the rebuttal deadline approaches, we would like to kindly follow up on our response and welcome any further feedback. If our clarifications address your concerns, we would greatly appreciate your reconsideration of the rating.

Thank you very much for your support!

Best,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair,

Thank you for considering our submission on SILK.

Key Strengths Highlighted by Reviewers

  • Practical relevance. oLjm highlighted the risk of cumulative knowledge leakage from individually innocuous queries.
  • Knowledge tracking and component ablations. Qnz9 and yb3r appreciated the informative ablations. yb3r further noted that atomic claims enable tracking of knowledge acquisition, while the ablations clarify each component's contribution to retrieval coverage and leakage.
  • Comprehensive evaluation and gains under defenses. Reviewers recognized our comparison with six baselines across four datasets, along with evaluations under alternative RAG configurations and multi-turn defenses. yb3r noted that SILK outperformed IKEA on all three metrics across all four datasets under both input- and output-side defenses.

How We Addressed Reviewer Concerns

  • Retrieval dependence and surrogate response agreement. We reran the original queries for SILK and all six baselines across four datasets using DeepSeek-V3 and GLM-4.7. With retrieval disabled or replaced by irrelevant context, both backends responded with “I don’t know,” yielding a leakage rate of 0.00 in both controls. Under standard retrieval, SILK consistently outperformed IKEA across all datasets and both backends (Tables 9–10\color{blue}{\text{Tables 9–10}}). These controls support that the measured leakage originates from RAG knowledge bases rather than model prior knowledge. Furthermore, disabling retrieval reduced the surrogate's ROUGE-L and semantic similarity scores to below 0.01 across all datasets, while the SILK surrogate outperformed IKEA's with retrieval enabled, supporting the contribution of extracted knowledge to response agreement (Table 17\color{blue}{\text{Table 17}}).

  • Claim repetition and document redundancy. After controlling for query round, partial rank correlations between claim repetition and observed retrieved-document redundancy ranged from 0.514 to 0.651 across four datasets (Table 11\color{blue}{\text{Table 11}}), supporting claim repetition as an observable proxy for document redundancy when retrieved documents are hidden.

  • Leakage metric validation and aggregation. Among five semantic cutoffs evaluated on 30 human-annotated pairs, 0.65 achieved the highest observed F1 (0.857) (Table 18\color{blue}{\text{Table 18}}). We reported the sample breakdown by dataset and method (Table 19\color{blue}{\text{Table 19}}), along with confusion counts, sample-level annotations, annotator agreement, and Cohen's κ calculations underlying Table 8\color{blue}{\text{Table 8}} (Appendix B.3, Tables 12–16\color{blue}{\text{Appendix B.3, Tables 12–16}}). We also clarified chunk deduplication and the computation and rounding of EE, LR, and SR (Appendix B.3 and Section 5.1\color{blue}{\text{Appendix B.3 and Section 5.1}}).

  • Defense evaluation and consistency across settings. We evaluated input- and output-side defenses, including single-turn and multi-turn detectors (Table 1 and Appendix F\color{blue}{\text{Table 1 and Appendix F}}). We clarified target/shadow detector configurations, chronological train/test splitting, and threshold calibration on 200 held-out benign dialogues, yielding a 4.5% query-level false-positive rate on calibration data (Appendix F.4–F.5\color{blue}{\text{Appendix F.4–F.5}}). SILK achieved higher success rates than IKEA in all 48 paired comparisons across RAG configurations, defenses, and datasets, with an average gain of 19.9 percentage points (Appendix D.4, Table 26\color{blue}{\text{Appendix D.4, Table 26}}).

We sincerely thank the Area Chair and reviewers for their careful evaluation and constructive feedback.

Best, 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.