acceptodds
Under review as a conference paper at ICLR 2027

M2AC: Policy-Aware Access Control for Multimodal Retrieval-Augmented Generation Systems

Abstract

To address the fine-grained access control problem for cross-modal data in multimodal Retrieval-Augmented Generation (RAG) systems, this paper proposes a policy-aware access control framework, M2AC (Multimodal Access Control). Centered on multimodal derived data such as text, images, OCR text, captions, summaries, and embedding vectors, this framework extends the traditional binary allow/deny decision into a policy control mechanism that covers the entire process of retrieval, reranking, fusion, summarization, citation, and generation. M2AC comprises two core components: the Derived Permission Propagation Engine (DPPE) and the Modality-Aware Policy Enforcer (MAPE). Experiments on a LangChain-based multimodal RAG with three vector databases, namely Milvus, Weaviate, and pgvector, show that M2AC's policy generation takes 48.3 seconds, on the same order of magnitude as recent RAG-specific access control methods while using the fewest rules; the F1 score of end-to-end policy enforcement reaches 96.8%, an improvement of 3.4 to 7.2 percentage points over these methods. Its average retrieval latency increases by only 4.7% compared to the lightest baseline, and its average defense success rate against four types of cross-modal attacks, such as OCR bypass, reaches 96.2%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
access control, multimodal retrieval-augmented generation, derived permission propagation, policy enforcement, cross-modal attack

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9CaMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Multimodal Access Control (M2AC), which assigns access control decision for cross-modal data in multimodal RAG systems. The proposed M2AC framework consists of two parts, Derived Permission Propagation Engine (DPPE) and Modality-Aware Policy Enforcer (MAPE). DDPE uses information retention ratio ρ and attenuation rule to measure the permission of nodes under different operations. MAPE assigns allow/degrade/deny permission for requests according to the DDPE rule. Empirical evaluations show that the proposed method produces the fewest rules at comparable generation time and has a high defense success rate against four types of cross-modal attacks.

Critical strengths

This paper proposes a complete modality-aware policy enforcement mechanism covering the entire process from retrieval to generation for multimodal RAG systems. It gives a new idea of how to achieve permission control for objects under operations (e.g., OCR, caption, summary, embed, crop,and translate) used in multimodal RAG systems. This is an interesting and timely contribution to the system access control area.

Critical weaknesses

There is no theoretical guarantee for the selection of each retention ratio ρ and threshold τ. They are selected empirically. The empirical evaluations show that the defense success rate does not reach 100% for all attacks. Therefore, the confidential objects in the system are not guaranteed to be secure and are at risk of leaking. I don't think this is acceptable for high-security systems.

Questions for authors

(1) Could the authors explain why the permission can be degraded after these mentioned operations? What is the advantage of this design? Does the advantage overweigh the additionally introduced risk of information leakage?

(2) Could the authors provide a detailed system scenario that M2AC can apply to? What is the security level of this system? Is the average defense success rate of 96.2% acceptable to this system?

(3) Since the threat model assumes an active attacker/insider holding low-privilege credentials. What is the defense success rate when the attacker combines all attack methods and use more adaptive attacks (e.g., reverse-engineer permission boundaries)? Although this is discussed as a limitation, it is very important to conduct such analysis for security concerns.

Reviewer_2b2oOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper looks at access control in multimodal RAG systems and proposes M2AC, which has two parts: Derived Permission Propagation Engine (DPPE) and Modality-Aware Policy Enforcer (MAPE). DPPE models derivation relationships as a DAG. Each edge carries an operation type and an attenuation function. MAPE enforces policies at six RAG stages using deny-override semantics. Policies are generated automatically in three steps: static and dynamic data-flow analysis, Ward hierarchical clustering of access logs, and a greedy set-cover minimization.

Critical strengths

S1: Multimodal RAG systems do generate numerous derivative artifacts, and privilege leakage resulting from OCR text, image captions, or summaries is a genuine issue for document-level filtering.

S2: Modeling derived outputs as a DAG with stage-level permissions is a natural approach.

Critical weaknesses

W1: The contribution is mainly a systems and security engineering design: a provenance DAG, ABAC rules, Casbin integration and greedy set cover. There is almost no learning component or representation-learning insight. Each idea has well-known precedents, the combination is sensible but conceptually incremental.

W2: The summary over-privilege defense depends partly on prompt engineering. This is known to be unreliable, and the paper gives no robustness evaluation against jailbreaks.

W3: The paper says "each attack was designed with 100 simulations" in line 366. With 100 trials per attack, success rates must be integers, yet Tables 5 and 6 report values such as 99.1%, 95.8%, 93.6%, 98.7%, 99.8% and 92.4%, including per-scenario numbers that would need even fewer trials. Either the protocol is described incorrectly or the numbers are not what they are claimed to be. No variances, confidence intervals or repeated runs are reported for any table.

W4: The end-to-end F1 score for policy enforcement requires a ground-truth label of whether each request should be allowed, denied or degraded. It is not explained who produced these labels, how degrade-vs-allow correctness is scored, or whether the labels come from the same policy specification M2AC is built from. If the ground truth comes from the same propagation rules, the comparison favors M2AC by construction. The same concern applies to the "Pass" metric.

Questions for authors

see the weaknesses

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

This paper proposes a policy-aware access control framework for multimodal RAG. The core idea is to move beyond binary allow and deny decisions and enforce access control throughout the multimodal RAG pipeline, including retrieval, reranking, fusion, summarization, citation, and generation. However, the paper appears to focus primarily on access control and security issues in multimodal database, with RAG serving mainly as an application scenario rather than the core technical object of study. Although the work presents some degree of novelty, its contribution is relatively limited and is more database security-oriented than AI-oriented. As a result, its fit with the central themes of AI research appears somewhat weak.

Critical strengths

The paper addresses a weakness in multimodal databases. Access control is often enforced only at the document or storage layer, while sensitive information may leak through derived artifacts such as OCR text, captions, summaries, embeddings, or citations.

Critical weaknesses
  1. The reported defense effectiveness appears to largely stem from M2AC’s explicit propagation of access policies to derived multimodal objects. The four evaluated attacks are all centered on accessing restricted information through derived representations, such as OCR text, captions, summaries, or retrieval-side signals. Since M2AC is specifically designed to assign and enforce permissions over these derived objects, its strong performance is expected. In contrast, most baselines do not explicitly model derived-data permission inheritance in multimodal RAG pipelines. Therefore, the experimental comparison may primarily reflect a difference in functional coverage rather than a direct security advantage under equivalent capabilities.

  2. The paper mainly studies how to assign permission levels to derived data according to automatically generated policies. Its novelty is relatively narrow and limited, and it does not sufficiently support the paper’s claim of providing an end-to-end security framework. From both the system design and the experimental evaluation, the work appears closer to an access-control study for multimodal databases, with RAG serving largely as a framing scenario. It does not substantially address the core security problems of RAG systems.

Questions for authors
  1. How are user-side permissions modeled and matched with data policies? It is unclear how user roles or attributes are represented, obtained, or synchronized from existing IAM, RBAC, and ABAC systems, and how they are compared with the effective policies of original or derived objects during authorization. Are the effective data policies assumed to be identical for all users, or do they vary according to user roles, attributes, and contextual conditions?

  2. How does M2AC constitute an end-to-end defense framework for RAG security? The introduction and related work mention prompt injection and knowledge-base poisoning as key RAG vulnerabilities, but the proposed method and experiments focus only on a limited set of cross-modal access-control bypass attacks. Can M2AC mitigate prompt injection or poisoning attacks, or should the security claim be narrowed to derived-data access-control attacks?

  3. Section 3.3 presents automatic policy generation at a relatively conceptual level. The authors describe three phases, namely derivation relationship discovery, access pattern learning, and policy rule optimization, but the concrete implementation details are underspecified. It is unclear how derivation edges are detected in static and runtime data flows, how access-pattern features are encoded and clustered, how clusters are converted into candidate allow, deny, or degrade rules, and how the greedy constrained set-cover algorithm resolves conflicts or prevents over-privilege.

  4. The threat model assumes that the attacker is an insider with legitimate low-privilege credentials who can issue normal RAG requests, but OCR bypass and caption leakage seem to require access to derived OCR and caption objects, cross-modal existence inference requires observable side channels such as result counts or scores, and summary over-privilege requires restricted content to enter the summarization context. Please clarify what the attacker can query, which tools or modalities can be invoked, what outputs are observable, and why these attacks are feasible under the stated threat model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.