acceptodds
Under review as a conference paper at ICLR 2027

Beyond Update Geometry: Post-Commitment Auditing of Individual Decision Influence for Proposal Admission in Decentralized Federated Graph-Level Clustering

Abstract

Decentralized federated graph-level clustering must admit peer proposals while preserving a shared clustering state. Updates that appear ordinary in parameter space can change assignments near small decision margins, so nominal authority need not reflect decision influence. To audit these effects while preserving execution continuity, we introduce Proof-Carrying Operator Consensus (PCOC). Peers commit aligned operator and claim pairs before an unlabeled audit challenge is sampled. Validators measure each proposal's leave-one-out effect and compute simultaneous upper bounds for individual influence in the frozen audit pool. The policy screens proposals using current and stored upper scores, then ranks passing proposals to select a fresh quorum. Stored pairs fill the remaining positions. Across five datasets with matched budgets, mean individual influence was 45.8% and 45.6% lower than for an internal reference under fixed and adaptive attacks. In a separate study with two encoders and 20 paired seeds, attack-induced boundary-flip rates (AI-BFR) were 6.3% and 5.2% lower, corresponding to approximately 0.052 and 0.047 percentage points. Neither difference passed Holm adjustment, with values of 0.0625 in both attacks. Final clean normalized mutual information (NMI) was 0.1199 versus 0.1217. A transfer record covering 50 contexts found lower influence in the execution pool for PCOC in 43 of 50 contexts.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Decentralized Federated Learning Graph-Level Clustering Decision Influence Auditing Proof-Carrying Operator Consensus (PCOC)

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_U6DoOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PCOC. After operators submit and commit their operator-declaration pairs, public beacons are used to extract "audit probes." The method calculates the leave-one-out impact of removing an individual proposal on the soft-clustering output, along with an empirical Bernstein upper bound. A deterministic gate and a pre-declaration sequencer are then employed to select a new quorum, thereby incorporating decision-making impacts, into the admission audit process. Experimental results across five graph datasets, under both fixed and adaptive attack scenarios and with matched budgets, demonstrate a significant performance advantage.

Critical strengths
  1. Following the commitment freeze, COC performs sampling to derive simultaneous upper bounds on the "leave-one-out" clustering impact of each proposal within the frozen audit pool. This approach captures the actual shifts in soft-clustering outputs that occur when a proposal is removed, addressing the limitations of methods like Krum or Trimmed Mean, which rely solely on parameter or update distances.

  2. Deterministic gates and pre-declared sequencers render the new quorum selection process both verifiable and reproducible.

Critical weaknesses

Proposition 1 only guarantees coverage for the "leave-one-out" impact of individual proposals within the audited pool; it does not formally cover more general counterfactual scenarios, such as the simultaneous removal of multiple proposals or the substitution of stored entries, which would require separate definition.

Proposition 2 decomposes the probability of an assignment flip into a "probe probability near the decision boundary" plus an impact term. However, obtaining an actual numerical bound still requires estimating the boundary probability component. There remains a gap between the theoretical framework and its numerical application, a point that warrants further clarification from the authors.

Questions for authors

Please see the weaknesses.

Reviewer_s5xJOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes PCOC, a commit-then-audit admission protocol for decentralized federated graph-level clustering that bounds each proposal’s leave-one-out influence on cluster assignments instead of screening by update geometry. It cuts measured influence by about 46% versus an internal reference.

Critical strengths

None

Critical weaknesses
  1. This paper is very hard to read, after reading Introduction 3-4 times, I still fail to understand the research problem, the motivation of this work, and the intuition behind the proposed solution. More specifically, you can almost find no clear connection among sentences within same paragraph, such as paragraph 3 of the introduction.

  2. The contributions are also extremely unclear, in the sense that they are not relative to any existing works, it is extremely hard to place this work in the landscape of the current research, if there is any.

  3. In Table 2, it appears that "Internal ref." brings the most significant performance gain over other baselines, however, "Internal ref." never got formally defined in the paper, and its relationship to the proposed method remains unclear.

Questions for authors

please refer to weaknesses

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

The manuscript introduces PCOC for decentralized federated graph-level clustering, accounting for the fact that updates that appear geometrically fine and acceptable, are indeed degrading the performance. To this end, graph encoders supply the representations on which the submitted cluster operators act, while aligned soft assignments make the effect of each operator observable on unlabeled graphs. The procedure commits the operator and claim components, freezes the audit pool, and samples a common task, from which simultaneous upper bounds on individual influence are obtained. A deterministic gate compares the predeclared current and stored scores, and a ranker selects a fresh quorum among the proposals that pass the gate.

Critical strengths

The mathematical framework is solid and the theoretical results (Props. 1 and 2) are stated clearly, positively contributing to the soundness of the manuscript. Moreover, the proofs for Propositions 1 and 2 in the Supplementary are easy to follow.

Critical weaknesses

I believe that the main weakness of the manuscript consists in contextualizing the theoretical framework to the actual scope of the paper. In particular, the authors did not clearly motivated their mathematical design choices, which makes the overall manuscript hard to follow.

Most of the symbols are used without a formal definition, and at times some are not even defined

The method, i.e., Section 3, is introduced ex-abrupto, without a proper framing and background; thus, making the paper not very accessible to a general FL reader who is not a specialist in federated clustering

Questions for authors

How does the limit f<n/3f < n/3 affect the algorithm? As far as my experience is concerned, most recent robust FL algorithms are designed to work without any bound on the fraction of malicious clients in the federation, for instance wavelet representation based method?

Could the authors clarify and justify the choice of the L1L^1 norm in line 200?

Could the authors discuss how their presented framework relates to privacy concerns?

Could the authors provide insights on the cost complexity of the method? How does efficiency improve or degrade?

Minor. In order to improve readability, I advise the authors to move Figure 2 after their method is introduced.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.