acceptodds
Under review as a conference paper at ICLR 2027

Priview: Trust-Source Decomposition for Robust and Composition-Safe Disclosure in Federated Multi-Agent Systems

Abstract

Federated multi-agent systems deploy agents at separate clients to collaborate through natural-language exchanges and multi-round negotiation. Each agent must contribute useful local information while respecting its client's privacy policy. Prior view-isolation methods extract abstractions of private information as views using trusted task descriptions or application interfaces, then restrict subsequent disclosure to those views. Open negotiation instead requires constructing such information from changing, untrusted requests and validating it together with prior disclosures, exposing the process to two key privacy threats: prompt-injection and differencing attacks. To address these threats, we propose Priview, a disclosure framework centered on trust-source decomposition. We first derive disclosure constraints from the integrity principle of robust declassification and the requirements of task utility and cumulative safety. Utility requires requests to guide information construction, whereas authorization integrity requires permission judgments to remain independent of untrusted request text. These requirements motivate trust-source decomposition: separating construction guided by low-integrity requests from authorization grounded in high-integrity, owner-controlled policies. Priview implements this decomposition in three isolated contexts: view construction proposes and ranks useful candidates, authorization checks them against the policy and disclosure history, and a confined speaker expresses approved content. A conservative ledger consolidates policy-relevant facts from actual disclosures for cumulative authorization. Across public benchmarks and the evaluated prompt-injection, differencing, and extended-context settings, Priview retains task utility with no observed unauthorized disclosures.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent, Federated, Integrity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6k9yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper introduces Priview, a framework for information disclosure in Federated Multi agent LLM based systems. Priview does it in three step process with each step having a clear role. V proposes candidate views from the request, A authorises with the help of owner policy and ledger and finally S gives the response using only the approved view. The framework formalizes this with robust declassification and measure performance in terms of utility and privacy. The paper reports near raw utility scores and zero items leaked on multiple datasets.

Critical strengths
  1. The problem is well motivated. Introduces a new angle to information disclosure by using robust declassification into LLM disclosure scenarios.
  2. Design of the solution for prompt injection and differencing attack looks good. A never sees the actual request sent and so no influence on the authorization outcome. S's only private input is the approved view, so it cannot reveal private data beyond it.
Critical weaknesses
  1. The zero leaks mostly comes from the structure and setup. Views are not generated(fixed for a scenario) and consent is given to Priview as strutured input. Also, auditor uses the same rules that the authorizer used. So, the set does not allow Priview leaking, but the baselines do not have the same setup. They have to infer consent from request itself.
  2. The auditor used might miss leaks when paraphrases appear, since it uses exact substring matching. The auditor also decides a message was authorized by a marker phrase inside the message text itself
  3. Weak Baselines in general. CoT+Sens is forced to disclose and ToM is not allowed to tell exact values.
Questions for authors
  1. Does the zero leak counts change if the auditor determines authorization from the round and fact identity instead of using a marker phrase in the message text. Also, will it change if use some kind of semantic check for catching leaks?
  2. Can the consent-aware 1-2-3 check looks like fair baseline comparison. Since the same permissions as Priview are given it would perform on similar levels is the expectation. Maybe that can used as a stronger baseline ?
  3. Can V generate views instead of taking the fixed views from the scenarios? Can it generate the view and rank them? How would it affect A's performance in that case
  4. ToM baseline tracks history (some kind of scratchpad to track states would be used) and also has zero combination leaks . What does the ledger add for privacy over general ToM methods?
  5. Can we measure the effectiveness of the compression of the ledger proposed? Is some kind of evaluation of the compression possible?
  6. Small Nitpicking, but can the model versions used for evaluation be mentioned more clearly?
3 comments

The agentic experiment retains task utility and zero detected leaks with generated views. In the supplementary agentic experiment, we replaced V’s fixed-candidate selection with LLM generation of 1–4 candidate disclosures from the request and local evidence, each with a predicted utility score. A then assesses semantic compliance and checks encoded policy rules deterministically. On DeepSeek HiddenBench, this implementation achieves 91.7% accuracy and 48/48 runs without detected leakage.

Baselines using the owner’s privacy policy retain Priview’s safety advantage in the cumulative-disclosure case. We removed forced disclosure from CoT+Sens and the extra exact-value restriction from ToM. Both now use the same owner policy and trusted permissions as Priview. On Qwen HiddenBench, CoT+Sens+consent, ToM, and Priview achieve 60.4%, 72.9%, and 70.8% accuracy. Their zero-leak counts are 42/48, 47/48, and 48/48 under deterministic auditing, and 30/48, 6/48, and 48/48 under semantic auditing, respectively. Check+consent also uses the owner’s privacy policy and matches Priview at 70.8% accuracy and 48/48 deterministic zero-leak runs. The difference between Check+consent and Priview lies in cumulative disclosure: in the case with 20 intervening discussion rounds, Check+consent leaks in 3/3 runs versus 0/3 for Priview.

Semantic auditing detects additional baseline leaks while Priview remains at 48/48 zero-leak runs. A separate DeepSeek post-hoc auditor checks paraphrases, entailment, and cross-round inference, taking a majority vote over three judgments per owner. The Qwen zero-leak counts change from 42/48 to 30/48 for CoT+Sens+consent and from 47/48 to 6/48 for ToM, Priview remains at 48/48. This identifies semantic leaks missed by fragment matching, including in the history-aware ToM baseline.

Priview supports round-specific permission activation without changing the original fact-based authorization conditions. The original datasets specify permissions by fact, so the round-based extension does not change their authorization conditions. We separately constructed a case in which permission takes effect in round 4: Priview withholds the protected information in rounds 1–3 and discloses it in round 4. Semantic fact auditing addresses a different question—what protected information the recipient can infer. As reported above, this stricter audit detects additional baseline leaks while Priview remains at 48/48 zero-leak runs.

The ledger reduces authorization-input characters by 96.8% in the stress case. In the 20-round stress case, supplying full history to A yields 2/3 leaks, versus 0/3 with the ledger. Mean authorization input falls from 10,574.7 to 337.0 characters, a 96.8% reduction. This directly measures the ledger's compact representation in addition to retaining history.

The revision will specify exact model versions and evaluation settings. The experiments use DeepSeek-v4-flash, Qwen-max, and GPT-5.6-Sol backbones. We will specify their exact model versions, run dates, and judge settings in the revised experimental setup.

Thank you authors for the detailed comment. The new experiments address most of my concerns and the check+consent comparison is a good addition. But, these results would replace much of the paper's evaluation, and they were run with small run counts, one benchmark family, and an LLM judge as the semantic auditor. So, I will retain my rating as my final score and will increase soundness score for the paper.

Thank you for recognizing the additional evidence and increasing the soundness score. We would like to clarify what the Check+consent comparison establishes.

Check+consent incorporates trusted owner-policy authorization, which is central to our trust-source separation principle. This addition gives the baseline the same owner-permitted information as Priview, rather than requiring it to infer permission from the request. Their comparable performance on short HiddenBench interactions shows that this authorization interface enables useful disclosure in both methods.

The remaining difference appears when individually permitted disclosures accumulate across rounds. The short benchmark interactions do not explicitly stress malicious requests or long disclosure gaps. We therefore added a targeted case with two complementary disclosures separated by 20 discussion rounds. Check+consent leaks the protected combination in 3/3 runs, while Priview has 0/3 leaks. This isolates the role of checking each proposed release against accumulated disclosure state, beyond providing trusted consent.

We will distinguish these targeted mechanism tests from the broader benchmark results in the revision.

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

This paper studies privacy-preserving disclosure in federated multi-agent collaboration, focusing on forged authorization and leakage from combining disclosures across rounds. Priview separates request-conditioned view selection from authorization based on owner-controlled policies and disclosure history, then confines the speaker to approved content. A structured ledger retains facts needed for cumulative authorization. The paper gives conditional safety arguments and evaluates the framework on decision-making, negotiation, and contextual-disclosure tasks, with ablations for authorization, view selection, and disclosure history. The reported results show that the framework can retain useful task information while preventing violations detected by the specified audits.

Critical strengths
  1. The experiments examine specific failure paths behind the proposed separation. In Table 3, making authorization request-visible leads to leakage in 20 of 28 forged-consent runs, whereas Priview has none under the reported audit. Table 11 also shows that providing the full transcript to view construction does not compensate for removing disclosure history from authorization. These controls provide useful evidence for maintaining a separate authorization boundary with persistent disclosure state.

  2. The framework preserves useful disclosures in the evaluated settings. On HiddenBench, Priview reaches 48/48 zero-leak runs on each backbone while matching raw accuracy on Qwen and ChatGPT. It also provides all 122 required ConVerse abstractions per tested model without exposing the underlying values. These results support the practical value of selecting permitted abstractions instead of relying on refusal.

Critical weaknesses
  1. The evaluation leaves an important part of the view-construction claim untested. Section 3.3 states that V ranks disclosures supplied by the scenario, such as summaries, buckets, and exact values. Selecting among these options does not require constructing an abstraction from private data in response to a new request. The results support useful selection and authorization over predefined views, but provide limited evidence for the construction capability that motivates the proposed departure from existing view-isolation methods.

  2. The HiddenBench utility comparison is difficult to interpret because the default 1-2-3 Check has a more restrictive disclosure rule. It can release only public-level candidates, while Priview selects among candidates approved under the owner’s policy. Its accuracy of 0.188 on every backbone therefore does not distinguish the effect of this restriction from the effectiveness of its reasoning. The consent-aware MAGPIE controls are helpful, but a corresponding HiddenBench comparison is needed to assess how much of the utility advantage persists when both methods can release the same owner-permitted information.

  3. The ledger ablations establish the importance of giving authorization access to prior disclosures, but do not isolate the benefit of the ledger representation itself. In the full-transcript control, V receives the complete conversation while A still checks candidates without disclosure history. This leaves open whether A would obtain comparable privacy and utility from the full factual history. That comparison would clarify what the ledger contributes beyond maintaining persistent authorization state, particularly given the paper’s emphasis on conservative compression.

  4. The privacy audit removes owner-permitted facts before checking whether protected information can be inferred, although those facts remain known to the recipient. An authorized annual mean, for example, can combine with a later partial-year mean to reveal a monthly value that the policy still protects. Removing the authorized mean can exclude this inference from the audit. The paper needs to clarify whether such consequences are intentionally covered by the original permission, since this determines the scope of the reported zero-leak results.

Questions for authors
  1. How are the scenario-provided candidate views constructed, and what task or policy information is used to prepare them? Can V produce an abstraction outside this predefined set? A worked example would help distinguish the information supplied by the scenario from the contribution of V.

  2. With V and S unchanged, how would authorization using the complete factual disclosure history compare with authorization using the ledger? Privacy, utility, and authorization-input size would help distinguish the ledger’s representational benefit from the benefit of retaining disclosure history.

  3. Suppose the owner permits an annual mean but continues to protect individual monthly values. Would a later disclosure of the first eleven months’ mean be rejected because the two reveal the final month’s value? Please clarify how authorization and auditing handle this case when Clean_P removes the permitted annual mean.

3 comments

The supplementary implementation generates request-specific views beyond the predefined candidates. For the supplementary agentic experiment, we replaced fixed-candidate selection with LLM generation: V constructs candidate texts, evidence combinations, and utility scores from the current request. A also uses LLM semantic assessment alongside deterministic policy checks. The scenario supplies only source facts and owner policies, and V decides how to combine and express the facts. The original implementation selected scenario-prepared disclosures at different granularities. The new implementation achieves 91.7% accuracy and 48/48 runs without detected leakage on DeepSeek HiddenBench.

Priview matches Check+consent on HiddenBench and prevents leakage in the cumulative-disclosure case. On Qwen HiddenBench, Check+consent and Priview receive the same trusted permissions and both achieve 70.8% accuracy and 48/48 deterministic zero-leak runs. Their difference appears in the cumulative-disclosure case: two individually permitted fragments, separated by 20 discussion rounds, jointly reveal a protected fact. Check+consent leaks in 3/3 runs and Priview in 0/3 runs.

The ledger reduces authorization input and prevents leakage in the full-history stress comparison. In the 20-round stress case, full-history authorization yields 2/3 leaks, versus 0/3 with the merged ledger. Mean authorization input decreases from 46.0 facts and 10,574.7 characters to 3.0 facts and 337.0 characters—a 96.8% character reduction. The intervening discussion includes irrelevant, repeated, and misleading authorization-related messages. The ledger retains policy-relevant disclosure state while reducing this distracting context.

The regression case verifies protection of monthly values despite an authorized annual mean. We added the annual/partial-year mean case and made changes: A retains cumulative checks when a candidate has explicit permission. Testing both disclosure orders verifies that Priview blocks disclosure of the protected monthly value through the combined means.

Reviewer_eaAEOpenReview ↗

Comment from Reviewer eaAE

Thank you for the additional experiments and the fix to the annual-mean example. These strengthen the paper. The remaining uncertainty for me is whether the original zero-leak results hold when the audit includes authorized disclosures as evidence for inferring other protected facts. This matters because the recipient still knows those disclosures. I appreciate the work on the rebuttal and retain my reservation about this part of the evaluation.

Thank you for clarifying this concern. We have re-audited the existing transcripts with authorized disclosures retained as evidence, and the zero-leak results remain unchanged in the evaluated sets. The revised audit finds no leakage in all 48 DeepSeek HiddenBench task–seed transcripts, and the DeepSeek MAGPIE results, including the forged-consent condition, also remain leak-free.

The revised audit retains authorized disclosures as inference premises and checks whether the resulting conclusions are protected by the policy. It flags a protected conclusion inferred either from an authorized fact combined with an unauthorized supplement or from two individually authorized disclosures. Thus, permission to disclose a premise does not exempt a different protected conclusion from the audit. These results directly address the concern about removing information that remains known to the recipient.

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

The paper tackles a federated multi-agent collaborative setting where no single agent has all the information to solve the joint problem and each has local privacy constraints. Agents receive requests and need to construct disclosures from their private data in way that contributes to solving the task (task utility) while satisfying the local privacy constraints. The paper considers two types of attacks: prompt injection attacks (on current request) and differencing attacks that exploit information disclosed through multiple rounds.

The paper’s approach is based on trust-source decomposition - which refers to the principle of separating the proposal of information needed for the task from the enforcement of constraints and policies. Priview separates request-guided utility estimation from policy grounded authorization.

The paper makes a systems contribution in converting this trust-source decomposition principle to a design separating view construction, authorization and speaker, and the validation is primarily empirical across benchmarks.

Critical strengths

The problem is an important one to tackle and was framed reasonably well, as were the overall design principles and high-level approach.

Critical weaknesses

The paper is lacking in details of how several components are implemented in practice. For example - how is the view construction V implemented and how is task utility predicted or estimated? How is the policy violation set constructed from the candidates? How is the speaker isolated from private information other than the candidates? Are all these steps / enforcement LLM-based or deterministic or hybrid? If so, and in particular for the authorization/cumulative safety is there a risk of errors (induced or otherwise)?

Questions for authors

Would suggest clarifying actual implementation.

1 comment

V and S use LLM generation, while A combines LLM assessment with deterministic policy checks. For the supplementary agentic experiments, we replaced fixed-candidate selection with LLM view generation and added LLM semantic assessment to A, retaining deterministic policy checks. We evaluated this implementation on DeepSeek HiddenBench (16 tasks × three seeds) and the cumulative-disclosure case. Its steps are:

  • V uses an LLM to generate and score views. Given the current request and local evidence, it generates 1–4 candidate texts with source IDs and predicts task usefulness on a 0–10 scale. Code validates the source IDs. Final task performance is measured separately using the HiddenBench answer or MAGPIE success criteria.
  • A combines LLM assessment with deterministic checks. It checks each candidate together with the disclosure ledger against the owner policy and trusted permissions. The LLM assesses semantic violations, and deterministic predicates check encoded policy rules. The violation set represents protected facts inferable from the candidate and prior disclosures.
  • S uses an isolated LLM call. Its input contains only the approved view and public context. Raw private data and rejected candidates are not visible to S, preventing direct access to that withheld information.

Priview defends against forged-authorization injection and checks cumulative disclosure safety. A bases authorization on the owner’s policy and trusted permissions, without receiving the current request, so claims of consent in the request cannot grant permission. A jointly checks the candidate and disclosure ledger, including previously authorized disclosures, to block combinations that reveal protected information. In the 20-round cumulative-disclosure case, Priview has 0/3 leaks versus 3/3 without the ledger.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.