acceptodds
Withdrawn from ICLR 2027

OCDR: Overlap-Aware Counterfactual Dual-Representation Scoring for Membership Inference in Retrieval-Augmented Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Retrieval-augmented generation (RAG) systems can leak whether a target sample is present in the external knowledge base. Existing membership-inference attacks can achieve useful ranking performance while still producing substantially overlapping member and non-member score distributions, leaving an ambiguous decision region. We study whether counterfactual reference evidence can be represented to improve score separability while retaining competitive attack utility. We propose OCDR (Overlap-Aware Counterfactual Dual-Representation), which represents each target using two complementary Out-RAG statistics: the raw deviation from the Out-RAG mean and its variance-normalized counterpart. The two features are standardized on development data, expanded with second-order interactions, and combined using an L2-regularized logistic scorer. On HealthCareMagic with Mistral-7B-Instruct-v0.2, OCDR achieves an AUC of 0.6614 and TPR@5%FPR of 0.260, compared with 0.6270 and 0.175 for DC-MIA. It also reduces KDE overlap from 0.9324 to 0.7743 and 80-bin histogram overlap from 0.895 to 0.570. Additional experiments across generator backbones and datasets further show that ranking performance, low-FPR detection, and score- distribution overlap capture complementary aspects of membership-inference be- havior. These results suggest that explicitly modeling and evaluating member/non-member score overlap provides a useful complementary perspective for membership infer- ence in RAG systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
OCDR: Overlap-Aware Counterfactual Dual-Representation Scoring for Membership Inference in Retrieval-Augmented Generation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_5eGEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing membership-inference attacks can achieve useful ranking performance while still producing substantially overlapping member and non-member score distributions, leaving an ambiguous decision region. This paper propose OCDR to sovle the above challenge exploiting two complementary Out-RAG statistics to calibrate the membership score.

Critical strengths
  • The paper studies an important and timely problem.
  • Extensive evaluation covering varying models and datasets.
Critical weaknesses
  • My main concern is the motivation of formulating member/non-member score separability as an explicit objective for RAG MIA. What advantages does this offer over existing MIAs which directly determine whether a sample exists in the external retrieval database?
  • The threat model seems unrealistic. Although the paper consider a black-box adversary, the reference RAG systems need to be match t he target system in several aspects, including system prompt, number of retrieved documents, the generat or and retriever. These assumptions seem overly strong.
Questions for authors
  • The paper consider a black-box access to the target RAG with substantial knowledge, including generator, retriever, number of retrieved documents and system prompt. Is this a practical configuration in real-world RAG deployments? A concrete, realistic instance can strengthen the paper.
  • The distinction of the paper from existing calibration MIAs is not clear. Its seems that this work is essentially a calibration, with differences only in its implementation.
  • I would consider raising my score if the authors adequately address this concern.
Reviewer_CxKdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors tackle MIA in RAG systems. They propose a technique called OCDR that combines raw and variance-normalized deviations from reference RAG responses using a quadratic logistic scorer for membership inference. Experiments show improved low-FPR detection on HealthCareMagic, but there are mixed results across datasets which questions whether it is an improved technique or not. The experimental comparison is limited to S-MIA and DC-MIA. Relevant attacks such as Interrogation Attack [1] and DCMI [2] are cited but not evaluated, while E-MIA [3] is not discussed. Comparisons under matched query budgets and clearly stated attacker assumptions would help establish OCDR’s contribution relative to these approaches.

[1] Naseh et al. “Riddle Me This! Stealthy Membership Inference for Retrieval-Augmented Generation.” ACM CCS, 2025.

[2] Gao et al. “DCMI: A Differential Calibration Membership Inference Attack Against Retrieval-Augmented Generation.” ACM CCS, 2025.

[3] Guan et al. “E-MIA: Exam-Style Black-Box Membership Inference Attacks against RAG Systems.” arXiv:2605.00955, 2026.

Critical strengths
  1. The compact scoring method improves low-FPR detection across various generators on provided HealthCareMagic dataset, demonstrating useful attack performance on that specific dataset.

  2. Reporting AUC, low-FPR detection, and distributional overlap provides a broader evaluation than ranking metrics alone.

Critical weaknesses
  1. The evidence for improved separability is inconsistent: Table 1 favors OCDR on Mistral KDE overlap, while Appendix D’s recomputation favors DC-MIA. Different development budgets for OCDR and DC-MIA further makes interpretation of gains unclear.

  2. The evaluation includes only S-MIA and DC-MIA. Interrogation Attack [1] and DCMI [2] are cited but not evaluated, while E-MIA [3] is not discussed. Comparisons with these methods under matched budgets and explicit attacker assumptions would better establish OCDR’s contribution.

Questions for authors

1.Which KDE computation should support the main conclusion, and how do you reconcile Table 1 with Appendix D?

  1. Does OCDR retain its gains when DC-MIA receives the same development budget? Also how does it compare to other prior works.
Reviewer_uayBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes OCDR (Overlap-Aware Counterfactual Dual-Representation), a membership inference attack against RAG systems. The paper argues that previous evaluation metrics, such as ROC-AUC, do not fully capture the overlap between member and non-member score distributions. To address this, OCDR constructs an additional feature (other than the raw deviation) from counterfactual Out-RAG references: The deviation in scores normalized by the standard deviation of Out-RAG scores.

Critical strengths

Unlike prior works, this paper highlights the importance of explicitly evaluating the overlap between member and non-member score distributions. This provides a complementary perspective on membership inference performance beyond ranking-based and low-FPR detection metrics.

Critical weaknesses
  1. It is unclear how distributional overlap is incorporated into the proposed objective. Although the authors introduce a distribution-based overlap coefficient (OVL) in Equation (3), its connection to the normalised deviation (f2) in Equation (4) is not sufficiently explained. In particular, why should normalising a sample-level deviation by the variability of Out-RAG scores help reduce the overlap between member and non-member score distributions? Furthermore, there is insufficient evidence that f2 provides additional information that is useful for membership inference. As shown in Table 2, adding f2 to f1 does not improve performance, and the OCDR model shows a noticeable advantage over the f1-only model mainly at TPR@5%FPR. This raises questions about the necessity of the dual representation and its connection to the stated goal of overlap reduction.

  2. The manuscript should be carefully checked against the ICLR 2027 Formatting Instructions, as several aspects appear inconsistent with the required format. For example, the instructions state: "Citations within the text should be based on the natbib package and include the authors’ last names and year (with the 'et al.' construct for more than two authors)."

  3. The code submitted by the authors contains the following statement in Chinese (I used an AI model to translate it, so I cannot guarantee the accuracy of the translation): "DC-MIA and OCDR were both evaluated on the same Holdout-A400 set. However, it would be incorrect to claim that DC-MIA was refitted on Dev600." According to the paper, OCDR was trained using Dev600. If DC-MIA was not trained or calibrated using the same development set, this raises concerns about the fairness of the experimental comparison.

Questions for authors
  1. Could the authors clarify how the normalised feature (f2) and the quadratic feature expansion are connected to the objective of reducing overlap between member and non-member score distributions? In particular, what justifies the expectation that these components improve distributional separability when overlap is not explicitly optimised during training?

  2. While f2 incorporates reference variability through normalisation and can therefore capture information not available from f1 alone, the current experimental evidence does not convincingly demonstrate that this additional information is useful (Table 2). Under what conditions does f2 contribute to improved membership inference performance? Could the authors provide qualitative examples illustrating cases where f2 leads to better predictions than f1 alone? More importantly, how substantial and consistent is this contribution?

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

Related papers

Discussion (0)

Sign in to comment.