acceptodds
Withdrawn from ICLR 2027

DuetER: Efficient Dual-Path Encrypted Search for Retrieval-Augmented Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Outsourcing retrieval-augmented generation (RAG) corpora and indexes to untrusted clouds creates a three-way conflict among retrieval fidelity, data confidentiality and scalability. Existing privacy-preserving RAG solutions either rely on single-path retrieval with inherently limited recall, or impose prohibitive cryptographic overhead. We propose DuetER, an efficient dual-path encrypted retrieval architecture that maps dense semantic similarity and compressed BM25 proxy onto a unified Euclidean ranking interface. Path-specific keyed indexes prune the search space to eliminate full‑corpus scans, the cloud ranks encrypted candidates, and the client recomputes native scores before lightweight query-adaptive fusion. DuetER designs compartment conditional approximate distance‑preserving encryption (\DPE), which enables efficient ciphertext ranking without directly exposing the single global comparison frame of conventional DPE. Independent aliases further remove direct identifier equality across paths. Experiments on real-world datasets show that DuetER outperforms both single-path variants and improves recall over state-of-the-art (SOTA) baselines. At million‑scale corpora, cloud‑side query processing completes in roughly 1 second. These results establish a measured utility–efficiency–leakage architecture for scalable dual-path RAG. Our code is available on https://anonymous.4open.science/r/DuetER-0D0E.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Retrieval-Augmented Generation, Encrypted Retrieval, Hybrid Lexical-Semantic Search, Conditional Approximate Distance-Preserving Encryption

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_f63LOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a new privacy-preserving retrieval system for outsourced RAG. The authors clearly state the threat model and argue that the proposed DuetER system delivers high nDCG comparing with multiple baseline methods.

Critical strengths
  • Privacy is clearly an important problem in RAG
  • I believe the authors attempt to cover all the important metrics and technical component of a security/privacy paper.
Critical weaknesses
  • Related work is overly sketchy. Two short paragraphs can hardly cover the important related works in this area.

  • The technical contribution over ppRAG seems incremental. It's unclear to me what is the significant new insight beyond ppRAG.

  • The presentation needs a lot of work, such as many overly short single-paragraph sections and unprofessional figures.

Questions for authors

Please articulate the technical challenge/gap between DuetER and ppRAG.

Reviewer_37UuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DuetER is a dual-path encrypted retrieval system for RAG corpora outsourced to an honest-but-curious cloud. It maps dense semantic similarity and a compressed BM25 proxy to a common Euclidean ranking interface. Residual spherical IVF and a keyed lexical posting index prune candidates, while compartment-local approximate distance-preserving representations enable server-side ranking. The client recomputes exact scores for returned candidates, deduplicates results through encrypted join capsules, and performs adaptive fusion.

The main contributions are this retrieval architecture, the compartmentalized coordinate-release mechanism, and an explicit leakage analysis with a leakage-relative wrapper security proof. Validation includes complete-corpus experiments on MS MARCO, Natural Questions, and HotpotQA, parameter ablations, comparisons with Pisces and PRAG, and selected inference attacks. The paper reports retaining over 98% of full-scan semantic nDCG@10 while scanning approximately 13% of corpus entries, with roughly one-second cloud processing on MS MARCO.

Critical strengths
  1. Evaluation at realistic corpus scale. The paper evaluates complete corpora containing millions of documents, reporting over 98% retention of full-scan semantic nDCG@10 while examining approximately 13% of corpus entries. The probe-count and return-depth ablations further characterize the quality–cost trade-off. This evidence supports the practical relevance of the architecture and its potential to make outsourced encrypted retrieval computationally affordable.

  2. Explicit treatment of leakage and guarantee scope. The paper documents numerical-coordinate and operational leakage, limits its simulation theorem to the cryptographic wrapper, and carefully restricts its identifier-separation claim. It also acknowledges that exact client-side reranking cannot recover discarded candidates. This disciplined accounting makes the proposed operating point easier to assess and provides a useful foundation for subsequent security analysis.

Critical weaknesses
  1. The substantive contribution is insufficiently isolated from established techniques. Dual-path private retrieval already exists in Pisces, while ppRAG provides a close lightweight distance-based alternative. However, the evaluation lacks an indexed ppRAG-style control and component comparisons that separate the benefits of indexing, compartmentalization, and learned fusion. Consequently, the reported gains do not establish which distinctive mechanism contributes substantial new value, leaving the significance of the work insufficiently supported.

  2. The central empirical privacy benefit is not adequately validated against attacks adapted to the construction. Appendix A.2 releases signed coordinate subsets of shared path-level intermediates, creating a potential cross-compartment alignment channel through repeated, jointly observed queries. Table 4 does not establish robustness against this channel or attacks combining coordinates, lexical postings, and operational traces. Because reduced geometric exposure is a central motivation for CCADPE, this gap materially weakens the claimed privacy–utility contribution. It concerns the consequences of the declared leakage, rather than a contradiction of the leakage-relative wrapper theorem.

Questions for authors
  1. Clarification of the security evaluation: Does Table 4 evaluate attacks adapted to the complete CCADPE server view, including cross-compartment coordinate alignment and combined posting/access leakage? Under Appendix A.2’s shared-intermediate construction, how does the evaluation account for correlations between coordinate differences across jointly observed queries?

  2. Evidence supporting contribution significance: What evidence isolates the benefit of CCADPE from candidate pruning and client-side fusion? In particular, how does it compare with a lightweight distance-based retrieval control using comparable indexes, candidate budgets, and fusion, with differences in leakage explicitly identified?

  3. Clarification of the evaluated pipeline: How does the lexical exact-reranking cap of 300 in Table 9 reconcile with the approximately 925–958 candidates described as client-reranked in Table 6? Which configuration produced the reported quality and latency results, and what rule selects the final exact-scoring subset across compartments?

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

This paper presents a solution for encrypted search that supports a hybrid of semantic (dense) and lexical (sparse) search. The authors propose Compartment CADPE on top of CADPE and normalize both dense and sparse search to Euclidean space to conduct an efficient cloud-based encrypted search supported by mean-squared residual spherical IVF index and collision-free keyed BM25 tail index, followed by client reranking and fusion. The authors claim that their solution achieves good balanced between privacy and efficiency while maintaining high utility.

Critical strengths
  • The authors notice the urgent need to support both dense and sparse search in RAG. They propose a few new primitives for supporting encrypted search, such as CCADPE and lexical path in a DPE pipeline. They successfully integrate other existing primitives into a working pipeline.
  • The authors provide empirical evidence for efficiency and utility. In particular, it retains most of the utility in plaintext search (hybrid with dense and sparse), while outperforming Pisces in communication overhead and online latency.
Critical weaknesses
  • The soundness of provided security analysis and missing evaluation of practical threats.
    • Thm 3 is close to tautological from my point of view. The leakage function LconcL^{conc} includes the full compartment coordinate views GcmpG^{cmp} (all document and query coordinates). The simulation proof therefore only shows that the PRF, PRP and AEAD wrappers add nothing beyond that. Everything that matters for confidentiality of embeddings and queries is inside the declared leakage, and that leakage is never bounded.
    • CCADPE is an affine map plus small bounded noise. The authors do not present a theoretical or empirical analysis of known-plaintext attacks (KPA), where with roughly d+1 known (plaintext, ciphertext) pairs in the same compartment, least squares recovers the map. The corpora used here are public. In a realistic deployment, many records could also be known or injected.
    • Compartments may be re-linkable into a global frame. All compartments of a path are coordinate subsets of the same globally rotated ciphertext, and it seems to me that the noise is sampled once for each document but not per copy. With rcr_c = 256 out of 512 semantic coordinates, two random subsets share about 128 coordinates in expectation. Each query sends coordinates to 128 probed cells, all derived from one uqu_q. Their union covers essentially all 512 coordinates. I believe an adversary could stitch replicas and query coordinates across cells. That would undo both the "no shared global frame" claim and the unrelated per-cell aliases.
    • Lexical index stores (alias,bl(i),Quant(xi,t))(alias, b_l(i), Quant(x_{i,t})) as posting entries. A lexical document has a single alias in its partition, and that same alias appears in every posting list containing the document. The server therefore holds a full document–keyed-term incidence matrix with quantized BM25 weights, plus document frequencies (BktSize), plus SearchEq for query terms. This is strictly more than classic SSE leakage, leading to concerns of leakage-abuse attacks (Islam-Kuzu-Kantarcioglu (NDSS'12) and Cash et al.'s count attack (CCS’15)).
    • The solution does not support malicious security, integrity, forward/backward privacy and updates.
  • Novelty of presented schemes: all components of the paper are largely known by the literature. For example, SAP/CADPE is adopted by ppRAG, ALSH MIPS-to-L2, feature hashing, mean-centered spherical IVF, HMAC-keyed inverted index (SSE), RRF, etc. The idea of supporting encrypted search on both dense and sparse index is also not new (Pisces); this proposed DuetER is faster indeed, but also more leaky in security. The compartment projection in CCADPE is indeed novel, but the security of such compartment raises concerns (see above).
  • Evaluations.
    • The 192.7x latency and 637x communication comparison in Table 2 is not like-for-like in that DuetER is more leaky by construction. Also, it uses only FiQA (57K documents, 300 queries), and the two systems ran on different machines (laptop vs. VM).
    • The leakage validation is shallow. It does not state an adversary with adaptive knowledge (knowing the deployed defense) and presents more powerful attacks.
    • RQ2 states "how does it compare against SOTA private-RAG baselines," but the comparison is limited to only 2 previous schemes, PRAG and Pisces. Authors are encouraged to present more baselines and compare more methods, either quantitatively or qualitatively. Potentially missing related works [1-7] and concurrent works [8-11]. In particular, MESS [8] assumes perfect cross-shard association in its formal bound. It also empirically evaluates linkage attacks. Tremblay Thibault et al. [9] argue that honest-but-curious encrypted protocols open reaction attacks. If the client acts on decrypted unverified data, its observable behavior becomes a function of the plaintext. DuetER’s Fetch step is exactly such a behavior.
  • Relatively unsatisfactory presentation. The figure 1 that presents the scheme is full of text and hard to follow. The caption is literally "Framework of DuetER" without any further elaboration. Additionally, the main body of the paper never mentions it at all. Authors are advised to revise this figure to contain better visual elements, more vivid icon layout, and use it to improve the introduction of their scheme, ideally with some toy examples to help general audience understand. The aesthetics of figure 2 and 3 can also be improved. Figure 2 looks compressed due to limited space, and figure 3 uses a 3D layout that introduces more confusion than what the information it offers.

[1] Yingfan Liu et al., “Privacy-Preserving Approximate Nearest Neighbor Search on High-Dimensional Data,” ICDE 2025. This is a DCPE-based vector search work.

[2] Jinhao Zhu et al., “Compass: Encrypted Semantic Search with High Accuracy,” OSDI 2025.

[3] Jingyu Li et al., “PANTHER: Private Approximate Nearest Neighbor Search in the Single Server Setting,” ACM CCS 2025.

[4] Sacha Servan-Schreiber et al., “Private Approximate Nearest Neighbor Search with Sublinear Communication,” IEEE S&P 2022.

[5] Alexandra Henzinger et al., “Private Web Search with Tiptoe,” SOSP 2023.

[6] Mingxun Zhou et al., “PACMANN: Efficient Private Approximate Nearest Neighbor Search,” ICLR 2025.

[7] Yulong Ming et al., “P²RAG: Privacy-Preserving RAG Service Supporting Arbitrary Top-k Retrieval,” arXiv (2603.14778).

[8] Haoyu Cui et al., “MESS: Fast and Private Semantic Search on Multi-Graph HNSW,” arXiv (2607.28999).

[9] Louis Tremblay Thibault et al., “Quantization Enables Private Dense Retrieval against Malicious Service Providers,” arXiv preprint 2026 (2609.36376).

[10] Peichun Hua et al., “Pointing the Way, Hiding the Destination: Practical Private Dense Retrieval at Scale,” arXiv preprint 2026 (2608.25735).

[11] Peichun Hua et al., “Spruce: Scalable Private Outsourced Retrieval Using Compact Embeddings,” arXiv preprint 2026 (2609.03376).

Questions for authors
  • Theorems 1–2 hold for a "fixed, key-oblivious workload," while queries are adaptive and IVF assignment is data-dependent. How does the union bound over M_{p,b} handle adaptivity?
  • Topic inference is 20-class. I suppose this suggests 20 Newsgroups (4,500 documents)? Are all Table 4 attacks ran on any larger deployed million-scale configuration with different (r_c, nprobe, replication r)? Leakage conclusions drawn from a 4.5K-document corpus may not transfer, and it's unclear how these choices can affect security.
  • Can you provide more like-to-like or qualitative comparisons and discussions to closely related works stated above?
Reviewer_M45GOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DuetER, an encrypted dual-path retrieval architecture for outsourced RAG. It maps normalized semantic embeddings and a compressed BM25 representation to a common Euclidean ranking interface, while retaining separate indexes and client-side fusion. Its main components include compartmentalized distance-preserving representations, path-specific candidate generation, and exact client-side reranking. Experiments demonstrate efficient retrieval over million-document corpora and evaluate several attacks against the exposed geometry.

Critical strengths
  1. The paper addresses a practically relevant problem. Supporting both semantic and lexical retrieval efficiently is valuable, and the integration of candidate pruning, cloud-side refinement, and client-side reranking forms a coherent architecture.
  2. The large-corpus evaluation is a strength. The semantic index scans approximately 13% of the vectors while retaining over 98% of the full-scan semantic nDCG@10. These results provide useful evidence that the system can substantially reduce search work while preserving retrieval quality.
Critical weaknesses
  1. Unclear security privacy and security setting. I could not find a clear explanation of what information may be revealed by the whole protocol. Please summarize what the data owner, cloud server, and authorized user initially know and additionally learn throughout the protocol. In particular, clarify which parties can access query information, document features, candidate identifiers, and retrieved contents, distinguishing intended outputs from permitted leakage and protected information.
  2. The baseline comparisons need strengthening and clearer qualification. The comparison with Pisces does not account for material differences in trust assumptions and permitted leakage. DuetER appears to trust both the knowledge-base owner and the authorized client and aims to protect the query from a third-party semi-honest cloud, subjected to the permissible geometric and operational leakage while Pisces addresses retrieval between a mutually distrustful knowledge-base holder and querying client. The reported speedup therefore confounds improvements in retrieval efficiency with savings obtained by relaxing security requirements. It does not demonstrate a more efficient solution under comparable protection. Meanwhile, the directly relevant ppRAG baseline is absent from the empirical comparison, despite both approaches using distance-preserving representations for encrypted semantic retrieval. An aligned comparison with ppRAG is essential to establish what CCADPE contributes beyond existing distance-preserving retrieval and the benefits of indexing and candidate pruning. The authors should explicitly compare representation construction, permitted leakage, retrieval quality, runtime, and communication under matched settings. Without this comparison, the evaluation leaves the central question of CCADPE’s incremental value unresolved. This is a substantive gap in the evidence supporting the paper’s contribution.
  3. The practical privacy benefits are not sufficiently validated against combined leakage. The security theorem explicitly permits coordinate and operational leakage, while the lexical index exposes label equality, posting incidence, posting sizes, and quantized weights. Table 4 does not make clear whether the evaluated attacks jointly exploit these signals alongside synchronized queries and access/fetch traces. This limits the evidence supporting the practical protection provided by compartmentalization. The evaluation should include a representative inference attack using the combined observable view and auxiliary information. The paper should also clarify whether query intermediates are reused across compartments and assess whether such reuse could enable cross-compartment alignment.
  4. The evaluation does not sufficiently separate the contributions of individual components. The parameter sensitivity studies do not establish how much each architectural component contributes to the reported gains. Targeted, controlled ablations are needed to distinguish the effects of candidate pruning, CCADPE refinement, client-side reranking, and DuetRank fusion. In particular, the full lexical pipeline should be compared against the same posting planner followed directly by exact client-side BM25 reranking, with quality, runtime, and communication reported under matched candidate budgets. The relationship between the 1,000-candidate planner cap, approximately 950 returned candidates, and the 300-candidate exact-reranking cap must also be explained. The measured online operations, DuetRank configuration, data splits, and uncertainty estimates should be specified to make the results reproducible and interpretable.
Questions for authors
  1. Can you provide an aligned semantic-retrieval comparison with ppRAG? For the Pisces comparison, which protocol steps and output functionality were benchmarked, and how do differences in trust assumptions and permitted leakage affect the interpretation of the reported speedup? Refer to details in weakness 1.
  2. How much does each major component contribute to retrieval quality and efficiency? Please provide targeted ablations, including the lexical posting planner followed directly by exact client-side BM25 reranking, under matched candidate budgets and evaluation splits.
  3. How are the approximately 950 returned lexical candidates reconciled with the 300-candidate exact-reranking cap? Please identify the selection rule and the candidate count associated with each measured stage, and clarify the timing scope, DuetRank evaluation protocol, and uncertainty estimates.
  4. Have you evaluated an attacker that combines compartment coordinates, lexical-index structure, synchronized queries, and access/fetch traces with auxiliary information? Are query intermediates reused across compartments, and could that reuse enable alignment between compartment views?

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

Related papers

Discussion (0)

Sign in to comment.