acceptodds
Withdrawn from ICLR 2027

RootCert: Certified RAG Under Bounded Authenticated-Principal Poisoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A single compromised writer can corrupt retrieval-augmented generation(RAG) by introducing numerous records into external knowledge sources, making record-count budgets sensitive to fragmentation and replication. RootCert addresses this issue by counting authenticated writer IDs, called principals, instead of records. It certifies a fixed query's reference answer against any finite append through at most principals after an immutable memory checkpoint. Our central contribution is a principal-to-selection reduction. We score records separately, convexly pool them into one bounded vector per principal, and select the top- ranked groups without merging them. If the identities, scores, and relative order of the unaffected candidates remain fixed, this deterministic procedure requires at most matched score replacements. The standard bounded-replacement argument then ensures answer invariance when the reference answer's margin over its strongest competitor exceeds . On StrategyQA, across two readers and three retrievers with frozen candidate exposure, the principal-based-certification maintains accuracy from to appended records, whereas the record-based scalar certification fails after six records. A RootCert-aware adaptive attacker alters of uncertified answers at without affecting any certified reference. Principal-splitting tests, additional readers, and HotpotQA are used to examine separately declared threat sets. Under these conditions, RootCert safeguards certified answers against any finite record volume within a principal budget.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-augmented generation, Certified robustness, Data poisoning, Authenticated principals, Robust retrieval and selection

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

RootCert limits each authenticated writer to one bounded contribution during answer aggregation. Under its selection assumptions, it certifies that an answer remains unchanged after appends from a bounded number of writers.

Critical strengths
  1. Counting writers addresses a practical weakness of record-based attack budgets. Splitting one malicious submission into many records does not increase the declared budget.
  2. The paper explains why grouping after passage-level retrieval can invalidate the guarantee. The stable-selection result connects the threat model to the final certificate.
  3. The replication experiment separates record volume from principal count. The adaptive attack also examines behavior outside the certified set.
Critical weaknesses
  1. RobustRAG [1] already establishes the isolate-then-aggregate framework. Could a simpler variant retain one record per principal before applying the same bounded aggregation rule? Compare this variant with convex pooling.
  2. The main experiments freeze the candidate buffer, while corpus-wide certification requires an exposure bound that holds for every allowed append. How can this bound be computed for a mutable retrieval index? Please demonstrate a full-corpus implementation that retains a useful certificate.
  3. The main experiment treats Wikipedia pages as immutable anchors and assigns synthetic IDs to attack records. Could the authors evaluate a corpus where benign records are also grouped by their actual contributors? This would test whether certification remains useful when many relevant passages come from one writer.

[1] Certifiably Robust RAG against Retrieval Corruption.

Questions for authors

See Weaknesses.

Reviewer_hzC5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a certification framework for RAG systems against poisoning attacks, which specifically considers the number of attacker-controlled principals rather than the number of poisoned documents (which is an approach that was mainly done in the previous work). To be more specific, the proposed approach groups documents by principal IDs, independently scores each document, and aggregates the scores into a single vector per principal. The authors provide a theoretical analysis showing that the number of affected principals bounds the number of score replacements under certain assumptions. They conduct experiments mainly on StrategyQA.

Critical strengths
  • I think the idea of providing robustness guarantees based on the number of attacker-controlled principals is interesting, although it relies on a relatively constrained setting with strong assumptions.

  • Also, while it is relatively highly constrained, the authors provide the theoretical analysis.

Critical weaknesses
  • The practical applicability seems to be limited. Specifically, rather than independently scoring each passage and aggregating the scores at the principal level (as assumed in this paper), many existing RAG systems retrieve documents at the passage level and jointly process multiple retrieved passages using an LLM when generating answers.

  • Similar to my previous concern, the proposed certification seems to be based on (relatively) strong assumptions (e.g., stable candidate exposure and principal-local processing), which may not hold in general RAG pipelines in a real world setting.

  • Isn't grouping documents by principal potentially limiting the utilization of useful information? In other words, the experimental setup does not really reflect realistic principle distribution. I am particularly concerned that the attack documents are assigned synthetic principle IDs. However, there could be a lot of different scenarios in the real world. For example, a single author (or principal) may contribute many many useful documents containing different and also complementary information, but all these documents are aggregated into a single vector.

  • Under the calibration settings in Table 2, the performance gap is much smaller than in the performances in Table 1. This seems to imply that the one of the biggest reason of the improved performance would be a calibration effect, rather than a proposed approach.

  • Can the proposed method (effectively) handle multi-hop reasoning across multiple documents? Many multi-hop questions require combining information distributed across different documents, but the proposed approach independently scores each document and aggregates the resulting score vectors.

  • Most experiments are conducted on StrategyQA, which consists of (relatively) simple Yes/No questions, while HotpotQA is only included as an additional generalization experiment (but with relatively low performance and it may not be the “best” dataset in this era, according to this post [1]). I highly recommend the authors to evaluate the proposed method on more diverse benchmark datasets, including recent QA datasets with more challenging multi-hop queries or open-ended answers.

[1] Why You Should Stop Using HotpotQA for AI Agents Evaluation in 2025.

Questions for authors

See above.

Reviewer_68bKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RootCert certifies RAG answers against append-only poisoning by budgeting the number of controlled authenticated writer IDs, or principals, rather than records. Each principal may append any finite number of records after an immutable checkpoint. The method pools record scores within each principal group and shows that, under principal-local processing and stable candidate exposure and selection, at most rr controlled principals induce at most rr matched selected-score replacements. A standard margin test then certifies answer invariance. Experiments primarily on StrategyQA examine certified accuracy, record replication, principal splitting, and adaptive attacks, with additional reader and HotpotQA evaluations.

Critical strengths
  1. Separates write authority from record volume. The principal budget allows one authenticated ID to submit any finite number of records. The paper identifies the conditions needed for grouping and top-KK selection to turn that budget into a bound on selected-score replacements, giving a clear formal treatment of record replication.

  2. Tests the boundaries of the guarantee. The experiments vary the number of records under one ID and the number of controlled IDs, include a RootCert-aware adaptive attack, and examine shared calibration and retrieval order. These results help distinguish passing a certificate, resisting the finite attacks tested, and actually answering correctly.

Critical weaknesses
  1. Limited comparison with related certification methods. The paper cites several relevant approaches, but the main text gives limited analysis of what the principal-to-selection reduction adds relative to them. The main experiments compare primarily with record-based scalar certification; the discussion of other methods does not clearly establish which conditions can be aligned or what relative benefit follows.

  2. Frozen candidates limit the demonstrated scope of certification. The one-replacement-per-principal bound requires unaffected candidate groups to remain available after an append. The primary experiments ensure this by freezing 100 checkpoint candidates per query. Section 4.6 shows that passage-first retrieval lets one principal's records displace multiple unaffected groups before grouping. Group-first retrieval avoids this in the tested appends, but a certificate for corpus-wide retrieval has not been established.

Questions for authors
  1. What conclusions does the principal-to-selection reduction add relative to cited certification methods such as RobustRAG? Which threat assumptions and certified outcomes can be aligned, and could you provide an experimental comparison for those cases?

  2. Does the exact group-first experiment in Section 4.6 reselect candidate groups from the entire corpus? For every allowed append, can it guarantee that unaffected groups remain exposed, or provide a uniform bound hh on additional replacements?

  3. Could you report the accuracy of using only the trusted prior p(q)p(q) in Equation (8), and the fraction of certified-correct decisions that would also be correct with p(q)p(q) alone? This would help establish how much the retrieved evidence contributes.

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

Related papers

Discussion (0)

Sign in to comment.