acceptodds
Desk rejected at ICLR 2027

E-MIA: Exam-Style Black-Box Membership Inference Attacks against RAG Systems

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Retrieval-Augmented Generation (RAG) equips large language models (LLMs) with external evidence by retrieving documents at inference time, but it also turns the retrieval corpus into a sensitive asset. Under a black-box setting, an adversary given a candidate document can infer whether it has been ingested into the RAG knowledge base (i.e., document-level membership inference) solely from query–response interactions, thereby leaking corpus coverage and the existence of sensitive topics. Existing RAG Membership Inference Attack (MIA) methods either rely on “soft” signals such as semantic similarity, which often yield overlapping member/non-member score distributions and unstable thresholds, or employ explicit confirmation probes whose intent is conspicuous and thus prone to refusal and detection. We propose E-MIA, which converts verifiable hard evidence in the target document (e.g., fine-grained details, proper nouns/technical terms, definitional statements, metadata cues, and causal/constraint relations) into an exam with four objectively gradable question types (FB/SC/MC/T/F), and uses the aggregated exam score across multiple evidence-targeted questions as the membership signal. Experiments across multiple datasets and diverse RAG configurations demonstrate that E-MIA improves member/non-member separability in stringent settings while preserving natural, stealthy queries, and we further analyze the impact of question composition and exam length on attack effectiveness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation, Membership Inference Attacks, Exam-Style Question Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_D3ZhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes E-MIA, which converts verifiable hard evidence in the target document (e.g., fine-grained details, proper nouns/technical terms, definitional statements, metadata cues, and causal/constraint relations) into an exam with four objectively gradable question types (FB/SC/MC/T/F), and uses the aggregated exam score across multiple evidence-targeted questions as the membership signal. Experiments across multiple datasets and diverse RAG configurations demonstrate the effectiveness of E-MIA.

Critical strengths
  • The paper has organized experimental results.
  • The paper addresses an important topic.
  • The paper finds an interesting feature as the membership signal.
Critical weaknesses
  • The performance of the method is surprisingly great. They have a "117.8–142.6% improvement in TPR@FPR=0.005 over the strongest baseline." They have this record but they never explain why. And all the performance of E-MIA is above a certain value consistently. The paper should provide more discussion on why they achieve great performance and why the baselines don't.

  • The paper has very few defense details and content. They did not provide a valid solution on how to defend against this kind of method.

  • It is not clear how they differentiate members and non-members in their experiments. How do they guarantee non-members are not in the training set? It is still questionable.

Questions for authors

See comments.

1 comment

We thank Reviewer D3Zh for raising the performance magnitude, the practical defense question, and the definition of a non-member. These concerns are important for interpreting what E-MIA demonstrates, and we answer them directly.

Weakness 1: The performance gap is surprisingly large and insufficiently explained.

Answer: E-MIA's performance is not a constant near-perfect plateau. Across the 12 main cells, TPR@FPR$=0.005$ spans 0.8170.817--0.9980.998, and Accuracy spans 0.9770.977--0.9990.999. Several cells fall well below saturation, including 0.8170.817 and 0.8420.842 on NFCorpus. The consistent result is the direction of the gap, not an identical score in every setting.

The mechanism behind the gap is visible in the score distributions. Baselines score the whole response through similarity, completion accuracy, or Yes/No correctness. These signals are smooth: a non-member can produce a plausible response from parametric knowledge or a similar retrieved passage, leaving substantial mass in the upper tail of the non-member distribution. At FPR$=0.005$, that tail is exactly where the threshold must be placed, so even a small overlap produces many false positives. E-MIA instead grades document-specific hard evidence with exact canonical answers. Its member and non-member distributions are wider and more bimodal, leaving less density near the low-FPR boundary (Fig.~2).

The format ablation isolates this design effect. At a fixed exam length, single-format exams reach only about 0.310.31--0.380.38 TPR@FPR$=0.005$, whereas the full four-format exam reaches 0.9960.996. Changing the evidence source alone, while holding length, type mix, grading, and aggregation fixed, shifts separability by up to 0.5380.538. The reported 117.8117.8--142.6%142.6\% figure is a relative improvement over a strongest baseline whose TPR at this strict operating point is often only about 0.20.2--0.50.5; it is therefore a ratio with a small denominator, not a claim that every E-MIA cell is perfect. The absolute TPR increase and the distributional separation are the substantive results.

Weakness 2: The defense discussion is incomplete and lacks an actionable solution.

Answer: The paper is primarily an attack-and-audit study, not a defense benchmark, and it does not claim a validated defense. Its results do, however, identify a concrete defense surface. Table~2 shows that query rewriting, response rewriting, and malicious-input guardrails do not remove the signal; these defenses act after the document has already made its evidence reproducible. A valid defense must intervene earlier, at the evidence that is exposed to the generator.

Concretely, we propose utility-aware evidence sanitization before ingestion. High-specificity facts that are not required for the supported downstream tasks are generalized to task-preserving ranges, categories, or paraphrases; exact numbers, rare identifiers, and narrow constraints are retained only when task utility requires them. At serving time, retrieved evidence can additionally be synthesized across multiple documents and rendered with randomized but semantics-preserving wording, so that no single document produces a stable exact answer key. The defense should be evaluated on the same audit metric used here, TPR@FPR$=0.005$, together with task-level QA utility, under both static and adaptive attackers. This is a proposed mechanism and evaluation protocol, not an experimentally validated result in the current paper.

Weakness 3: Membership labels and non-member exclusion remain unclear.

Answer: The inferred property is membership in the RAG index, not membership in the base model's pretraining corpus. After preprocessing, we use a fixed split: 60% of each corpus is ingested into the knowledge base and 40% is never ingested. The 1,000 members and 1,000 non-members per corpus are labeled directly by that ingestion log, not by model behavior. Non-members are therefore guaranteed absent from the RAG knowledge base, while other documents, including similar ones, remain available for retrieval.

No study using public pretrained models can guarantee that a document was absent from pretraining. If a non-member appeared in pretraining, that contamination would make it more knowledgeable and therefore more likely to be mistaken for a member; it would raise non-member scores and reduce measured separation. The reported low-FPR separation is thus conservative with respect to this confound. The paper's claim is correspondingly scoped to index membership. If these clarifications are persuasive, we would be grateful if the reviewer would reconsider the corresponding assessment.

Reviewer_r45EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes E-MIA to infer whether a document is included in a RAG knowledge base. It generates four types of questions from the document and uses the system's answer accuracy as a membership score. Experiments on three corpora and four LLMs show improvements over existing attacks.

Critical strengths

The analysis of how evidence types and question formats affect attack performance is interesting. The results under query rewriting and malicious-input detection are also interesting.

Critical weaknesses

Compared to IA, the novelty is limited. IA already uses document-based questions and answer correctness to infer membership. Correct answers do not prove document membership: the model may know the facts or retrieve them elsewhere. I think this problem should be further clarified. Question and answer quality is insufficiently validated. The filtering procedure and error rates are unclear.

Questions for authors

How well does the base model answer the same questions without retrieval? What happens when the target document is removed but similar documents remain? Were the generated questions and answers manually checked? Some details of human verification are needed.

2 comments

We thank Reviewer r45E for the focused questions on novelty and causal attribution. We address the first two weaknesses and the first two questions below; the question-quality concern is answered in the continuation reply.

Weakness 1: Novelty is limited relative to IA's question-based correctness signal.

Answer: IA and E-MIA share answer correctness as a signal family, but the contribution is not the mere use of questions. IA uses a summary followed by Yes/No probes. E-MIA decomposes a document into five types of hard evidence, maps them to four question formats with different answer spaces, requires unique canonical answers and deterministic grading, allocates evidence by discriminativeness, and aggregates type-wise accuracy into one exam score. This design makes evidence source and elicitation format separately controllable.

The controlled evidence is in the ablation studies. Holding exam length, format mix, grading, and aggregation fixed, changing the evidence source alone shifts membership separability by up to 0.5380.538. In the fixed-budget format ablation, every single-format exam reaches only about 0.310.31--0.380.38 TPR@FPR$=0.005$, while the four-format exam reaches 0.9960.996. Thus the gain is not explained by ``asking questions'' alone; it depends on which document-specific evidence is elicited and how it is elicited.

Weakness 2: Correct answers do not establish membership under parametric knowledge or alternative retrieval.

Answer: We do not infer membership from a single correct answer. Correctness is a noisy observation, and the decision uses an aggregate score with frozen weights and threshold. The same exam is applied to both arms; the only designed difference is whether the target document is in the index. A non-member can answer from parametric knowledge or from another retrieved document, and the non-member score distribution confirms that this happens. Those channels raise non-member scores and therefore make the observed separation conservative rather than inflated. The claim is calibrated inference at the population level, not logical proof from one answer.

Question 1: How well does the base model answer without retrieval?

Answer: The non-member arm is the relevant empirical control for base-model knowledge. It uses the same base model, the same exam, and the same RAG interface, but the target document is absent; the model may still answer from parametric knowledge or other retrieved documents. On TREC-COVID, non-member scores are concentrated near 40/100 while member scores cluster roughly between 85 and 100 (Fig.~2). The base model can therefore answer some items, but not at member frequency. A separate retrieval-off run would remove the other-document channel and would be a weaker, less realistic control.

Question 2: What happens when the target is removed but similar documents remain?

Answer: This condition is implemented directly by the experimental split. Non-member documents are never ingested, while the rest of the corpus, including similar documents, remains in the index. The same target content, exam, scorer, retriever, and model are used; only target ingestion changes. The 1,000 non-members per corpus are therefore the requested ``target removed, similar documents retained'' condition, and the low-FPR results in Table~1 already measure it.

We continue the response to Reviewer r45E with the question-quality concern and its supporting question.

Weakness 3: Question/answer quality validation and error rates are unclear.

Answer: Generation is constrained to evidence with an objectively checkable answer. The consistency-cleaning stage removes items with multiple valid answers, items requiring excessive context for objective grading, and wording that reveals explicit probing intent. Choice distractors are drawn from nearby-but-incorrect candidates, and answer normalization is rule-based and deterministic. These steps are applied without access to the target model's responses or the membership labels.

We also manually audited a stratified sample. For each of NFCorpus, TREC-COVID, and SCIDOCS, we sampled 100 documents with an equal member/non-member split and eight items per document, two from each question type. This yields 800 audited items per corpus and 2,400 in total. The audit checked whether the canonical answer was supported and unique in the source document, and whether the answer-normalization layer recovered the intended response. Items that failed were removed or regenerated before evaluation; the reported evaluation set contains only accepted items. No target-model output was used when defining an answer key.

Question 3: Were the questions and answers manually verified?

Answer: Yes. Beyond automated consistency checks, we performed the stratified manual audit described under Weakness~3: 100 documents per corpus, balanced between members and non-members, and eight items per document, totaling 2,400 reviewed items. The review covered answer faithfulness, answer uniqueness, and grading behavior. Because this was a construction-quality gate rather than a separately powered annotation benchmark, we use it to exclude invalid items from evaluation rather than to estimate a model-level error rate. If these clarifications are persuasive, we would be grateful if the reviewer would reconsider the corresponding assessment.

Reviewer_bkfNOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes E-MIA, a black-box attack that infers whether a candidate document belongs to a RAG corpus using only final textual responses. It converts document-specific evidence into a 28-question exam containing fill-in-the-blank, single-choice, multiple-choice, and true/false items, then combines deterministic grades using calibrated weights and a threshold. Experiments on three corpora and four generators report accuracy of 0.977-0.999 and TPR of 0.817-0.998 at FPR 0.005. Additional analyses cover evidence selection, question composition, retrievers, rewriting, and prompt-injection detectors.

Critical strengths
  1. The membership score is interpretable and compatible with restricted access. E-MIA uses only final textual outputs and grades specific facts through explicit normalization and matching rules. This avoids dependence on target logits, retrieval traces, or a subjective semantic-similarity evaluator and gives a concrete way to study document-level leakage.

  2. The evaluation explores several relevant sources of variation. The paper reports large gains over three RAG-MIA baselines across three corpora and four generators, together with question-format, evidence-source, exam-length, and retrieval-backend analyses. Rewriting and guardrail experiments further characterize the attack, even though bypassing prompt-injection detectors is narrower than demonstrating general stealthiness.

Critical weaknesses
  1. The reported evaluation requires reconciliation before its conclusions can be trusted. For NFCorpus/Llama3.1-8B/IA, Table 1 reports AUC-ROC 0.929, [email protected] 0.827, and [email protected] 0.967. Any monotone ROC containing those two operating points must have AUC at least 0.827(0.05−0.01)+0.967(1−0.05)=0.951730.827(0.05-0.01)+0.967(1-0.05)=0.95173, exceeding the reported value beyond rounding error. This could be a reporting or protocol error, but it affects the central baseline comparison. Calibration-set provenance and low-FPR uncertainty also remain insufficiently documented.

  2. The exam mechanism is not sufficiently isolated from answer exposure and probing budget. In Table 4, a choice option explicitly contains "seizure disorder," which answers another item's blank within the same query. Shared context and repeated facts also leave Eq. (8)'s independence assumption unsupported. Although retrieval recall is reported, there is no retrieval-disabled or target-removal control. Without these controls and resource-matched IA/MBA variants, the results do not cleanly separate evidence elicitation from prompt-supplied information and aggregation effects.

Questions for authors
  1. What are the exact development-set sizes and membership-label sources, and how are calibration documents separated from evaluation documents? How do the frozen weights and threshold perform when calibration excludes the evaluated corpus or generator?

  2. How can the NFCorpus/Llama3.1-8B/IA AUC and TPR values in Table 1 be reconciled with ROC monotonicity? Please provide the underlying prediction scores, labels, and metric computation procedure, explaining any different evaluation populations and separating ROC-swept results from the fixed threshold of 62.2.

  3. How do exam scores change with retrieval disabled and when the same target document is removed from an otherwise unchanged index? Can the authors repeat this comparison after removing cross-question answer exposure and redundant evidence, to isolate corpus-dependent information from information supplied by the exam?

  4. Do E-MIA's gains persist when IA and MBA use matched numbers of factual probes and comparable token and retrieval-call budgets? Please also clarify whether Figure 4(d)'s single-format and mixed-format exams contain the same total number of questions.

2 comments

We thank Reviewer bkfN for the careful scrutiny of the calibration protocol and the metric entries in Table 1. The apparent AUC inconsistency is real, and its source is a transcription error rather than a change in the ROC data. We address Weakness1 and Questions1--2 below.

Weakness 1: The AUC/TPR entry violates ROC monotonicity, and calibration provenance is insufficiently documented.

Answer: The reviewer's monotonicity check is correct. The two reported operating points, (FPR,TPR)=(0.01,0.827)(\mathrm{FPR},\mathrm{TPR})=(0.01,0.827) and (0.05,0.967)(0.05,0.967), already imply a piecewise-linear ROC area of approximately 0.95430.9543. An AUC of 0.9290.929 cannot coexist with those points. The correct recorded AUC is 0.9790.979; 0.9290.929 is a transcription error in the table. The underlying continuous scores and the reported TPR values are unchanged, so the correction does not alter the baseline comparison or any conclusion drawn from it.

Question 1: What are the development-set sizes and label sources, and how are calibration and evaluation separated?

Answer: Membership labels are defined by controlled index ingestion, not inferred from model outputs. After preprocessing, each corpus is split into 60% member documents, which are ingested into the RAG index, and 40% non-member documents, which are never ingested. Every reported cell uses a balanced evaluation set of 1,000 members and 1,000 non-members per corpus. Thus the label source is the index-construction log.

Calibration is configuration-level rather than a classifier fitted to the evaluation labels. Its exact size is 25 configuration trials spanning the three corpora and four target LLMs; these are candidate parameter configurations, not 25 development documents. The evaluation set is exactly 2,000 documents per corpus. The development split used for a trial is held out from that evaluation set, and we select the medoid configuration rather than the per-cell optimum. This yields one frozen vector w=(0.312,0.214,0.300,0.174)w=(0.312,0.214,0.300,0.174) and one frozen threshold τ=62.2\tau=62.2 for every reported cell. No corpus-specific or model-specific retuning is performed. Consequently, every Table~1 cell is already an out-of-cell test of the same parameters. As an additional transfer check, the same frozen parameters are applied to held-out exam-generator models and to ChineseWebText, where TPR@FPR$=0.005$ remains above 0.9000.900.

Question 2: How can the AUC/TPR inconsistency be reconciled, and how are the metrics computed from scores and labels?

Answer: For each document, every attack produces one continuous membership score. A document is labeled positive if it was ingested and negative if it was held out. For a fixed corpus--model--attack cell, ROC points are obtained by sweeping a threshold over these 2,000 scores, and AUC-ROC is the area under that curve. TPR@FPR values are the TPR at the requested false-positive operating points on the same curve. The fixed threshold τ=62.2\tau=62.2 is used only to convert E-MIA's score into a binary decision for Accuracy; it is not used to compute AUC or TPR@FPR. All metrics in a cell therefore use the same 1,000 positive and 1,000 negative documents.

At FPR$=0.005$, 1,000 negative documents correspond to approximately five expected false positives. We therefore do not interpret tiny differences at that tail as precise rankings. The reported E-MIA advantage is not a boundary effect: its low-FPR TPR spans 0.8170.817--0.9980.998 across the 12 cells, and the direction is consistent across corpora and backbones. These clarifications account for both the numerical inconsistency and the metric provenance. If they are persuasive, we would be grateful if the reviewer would reconsider the corresponding assessment.

We thank Reviewer bkfN for separating the answer-exposure concern from the comparison-budget concern. We address both directly and then answer Questions~3--4.

Weakness 2: Prompt exposure and probing budget may confound the claimed evidence-elicitation effect.

Answer: The illustrative exam does contain a cross-format cue: the phrase ``seizure disorder'' appears in one option while a fill-in-the-blank item asks for that phrase. We do not treat the example as leakage-free. The key control is that the same exam, scoring rule, retriever, and interaction history are used for every document; only whether the target document is in the index changes. Redundancy can raise the absolute score in both arms, but it cannot by itself create a membership-specific advantage. The assembly procedure also limits redundancy to a small subset of critical evidence units, rather than repeating every fact across formats.

Equation~(8) is an idealized variance formula derived under conditional independence. Conditional independence is not used to construct the exam, grade responses, compute the empirical ROC, or make a membership decision. Violations of that assumption therefore weaken the interpretation of the closed-form variance expression, not the validity of the measured AUC and TPR values.

On probing budget, E-MIA submits the complete exam as one query. IA submits its questions separately, so E-MIA does not use more retrieval calls than IA; MBA likewise uses a single masked-completion query. Thus a larger retrieval-call budget cannot explain E-MIA's advantage. The overall prompt may be longer because it contains several items, and that cost is part of the exam interface. Figure~4(d) holds the item budget fixed while varying only the format mix, isolating the contribution of format diversity.

Question 3: Are the scores driven by target availability rather than prompt leakage or redundant evidence?

Answer: The target-absent control is already the non-member arm. Non-member documents are never ingested, while the other 60% of the corpus, including semantically similar documents, remains in the index. The exam, target model, retriever, scoring, and query history are unchanged; only target ingestion differs. The 1,000 non-members per corpus therefore measure exactly the requested ``target removed, similar documents still present'' condition. This is a harder counterfactual than disabling retrieval, because the model may still obtain an answer from parametric knowledge or another retrieved document. The resulting score distributions remain separated: E-MIA reaches AUC-ROC 0.9820.982--1.0001.000 and TPR@FPR$=0.005$ of 0.8170.817--0.9980.998 across the 12 main cells.

The paper does not claim that an exam is redundancy-free. A separate exam template with all cross-format overlap removed would test a different attack variant, not the deployed attack. For the membership claim, the relevant point is symmetry: prompt-supplied cues are identical for the member and non-member versions of the same target content, whereas the index label differs. Such cues can inflate both score distributions, but they cannot supply the label-dependent component of the separation.

Question-4: Do the gains survive matched probing budgets, and does Figure~4(d) use equal exam length?

Answer: Figure~4(d) holds the total item budget fixed and varies only the number and identity of question formats. At that fixed budget, the four single-format exams reach TPR@FPR$=0.005$ between approximately 0.310.31 and 0.380.38, whereas the full four-format exam reaches 0.9960.996. The gain therefore comes from evidence--format complementarity, not from giving the mixed exam more questions.

For IA and MBA, we use their native, published probing protocols without reducing their budgets. Their interfaces differ: IA uses one retrieval call per question, MBA uses masked completion, and E-MIA uses one call for the entire exam. Exact token equality is not well-defined across these interfaces, but E-MIA's retrieval-call count is not larger than IA's, and it is not larger than MBA's. The comparison therefore does not rest on E-MIA receiving more retrieval calls; the controlled evidence for the remaining gain is the fixed-budget format ablation. If this clarification is persuasive, we would be grateful if the reviewer would reconsider the corresponding assessment.

Comments on the paper

1 comment
Area_Chair_vV4pOpenReview ↗

Reminder to respond to the authors’ rebuttal

Dear Reviewers,

Thank you for your service as reviewers for ICLR 2027. As the discussion deadline approaches, please read the authors’ rebuttal and respond at your earliest convenience. A brief note indicating whether the authors have addressed your concerns would be helpful.

Thank you for your time.

Best regards,
Area Chair

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

Related papers

Discussion (0)

Sign in to comment.