acceptodds
Under review as a conference paper at ICLR 2027

Characterizing the Topological Distribution of Compromised Clients in Decentralized Federated Learning via Model Deviation

Abstract

Decentralized federated learning enables interconnected clients to collaboratively train the machine learning model over a communication graph. Since clients distributed across different positions in the graph propagate information along different trajectories, the attack performance of compromised clients depends on both the attack strategy and the topological distribution of compromised clients (TDCC). However, most existing studies evaluate attacks by selecting compromised clients at random, thereby overlooking the heterogeneous ability of different clients to propagate malicious information. In this paper, we investigate how TDCC affects attack performance in a DFL system and propose Maximum Client-Level Accumulated Model Deviation (Max-CL-AMD), a compromised client selection method based on the system's mixing matrices. Specifically, we model the system-level accumulated model deviation (AMD), which captures the overall deviation induced by compromised clients, and decompose it into client-level AMD, each quantifying the deviation attributable to an individual compromised client. Extensive experiments across diverse settings show that Max-CL-AMD can achieve varying degrees of improvement in attack performance over random selection and other baseline selection methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Decentralized Federated Learning

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

The paper studies how the placement of compromised clients affects poisoning attacks in decentralized federated learning. It expands the learning dynamics through products of mixing matrices, defines system-level accumulated model deviation, and introduces a client-level score based on accumulated column sums. Max-CL-AMD selects the highest-scoring clients. Since the actual mixing matrices may depend on the attack and aggregation rule, the implemented selector uses a fixed topology-derived matrix, (D+I)^(-1)(A+I). Experiments compare placement strategies across graph topologies, aggregation rules, datasets, and data distributions, primarily using Gaussian perturbations in a 20-client system. The results show that placement can substantially affect attack performance, although the proposed selector often matches degree or PageRank selection.

Critical strengths
  1. The paper highlights a consequential evaluation issue: random compromised-client placement can underestimate attack effectiveness in decentralized learning. For example, Table 1 reports an increase from 18.52% TER under random placement to 65.34% under Max-CL-AMD for ER/Multi-Krum. This supports treating attacker placement as an explicit part of robustness evaluation.

  2. The evaluation includes several relevant comparators and settings, including degree, PageRank, SWAP-BPI, multiple robust aggregation rules, and multiple graph families. Reporting standard deviations and the selected-set overlap in Appendix F.4.4 is useful: it exposes both sensitivity to graph realizations and the substantial agreement between the proposed method and simpler selectors.

Critical weaknesses
  1. The central optimality claim does not follow from the bound or its proof. Appendix E, Eq. (30), establishes only that selecting the largest fixed scores maximizes their sum. However, the scores are omega_k(C,T): changing C can change the mixing matrices and hence every score. More fundamentally, maximizing an upper bound does not establish maximization of the actual deviation in Eq. (5), which includes perturbation directions, gradient feedback, and interactions. There is also a concrete gap between Eqs. (7) and (30): Eq. (7) sums over all N clients, whereas the proof optimizes a sum over C. Under nonnegative row-stochastic mixing, sum_k ||Phi(C,T,t)[:,k]||_1 = |N|, so the displayed all-client bound is independent of C. It therefore cannot itself distinguish placements. Appendix D additionally applies Assumption 5 to the difference between attacked and clean gradients of the same client, although that assumption is stated for differences between clients' gradients; the required cross-trajectory bound is not established. These are substantive gaps in the claimed theoretical characterization, not merely missing proof details.

  2. The implemented method's contribution beyond existing centrality heuristics, especially for nonlinear aggregation, is not convincingly established. Equations (10)-(11) use only the communication graph and horizon, so the actual selector does not adapt to the aggregation rule or data heterogeneity. For an undirected connected graph, the stationary distribution of this estimated random walk is proportional to d_k+1. Consequently, at sufficiently long horizons its leading accumulated-column-sum term ranks clients by degree, subject to finite-horizon corrections. The manuscript should analyze this relationship. Empirically, the proposed method matches PageRank on every ER and BA entry of Table 1 and almost all entries of Table 2. Even at 1,000 clients, Appendix F.4.4 reports identical proposed/PageRank selections for all BA seeds and four of five ER seeds; that experiment reports overlap, not large-scale attack performance. The evidence therefore supports the importance of placement more strongly than a distinct advantage of Max-CL-AMD. Moreover, the main Gaussian attack has unbounded support, whereas the analysis assumes a deterministic perturbation bound, and several results have very large standard deviations. A clear account of any clipping and paired comparisons where the selected sets actually differ are needed to assess the claimed improvement.

Questions for authors
  1. Can you provide a corrected theorem that explicitly states the optimized objective and handles the dependence of omega_k(C,T) on C? In particular, how do you obtain a C-restricted optimization from the all-client bound in Eq. (7), whose right-hand side is constant under row stochasticity? If the result only optimizes a fixed estimated surrogate, would you revise the optimality claim accordingly?

  2. What assumption justifies Eq. (25)'s bound on the attacked-versus-clean gradient difference? Also, how does one common scalar |N| x |N| mixing matrix represent coordinate-wise median or trimmed mean, where different coordinates can select different neighbors? Please specify whether the analysis instead requires coordinate-dependent weights or a different representation, and how the subsequent norm bounds change.

  3. What finite-horizon information does Eq. (11) capture beyond degree and PageRank? Please report the horizon used for selection and paired per-seed attack-performance differences for settings in which the methods choose different sets, including actual training results for the 1,000-client setting. Such evidence would materially affect my assessment of the method's added value.

  4. Were Gaussian perturbations clipped to satisfy Assumption 3? Please specify the learning rate, training/selection horizon, graph-generation parameters, robust-aggregator parameters and local Byzantine-count assumptions, and the source of randomness varied across the five runs. Clean no-attack TER and paired uncertainty estimates would help determine whether small reported gains reflect attack effects rather than variation in training or graph realization.

Reviewer_Ur1cOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how the placement of compromised clients affects attacks on decentralized federated learning. It expresses the difference between attacked and clean model trajectories through products of mixing matrices, derives a client-level accumulated-deviation score, and proposes selecting the highest-scoring clients. Since the actual mixing matrices are unavailable before client selection, the implemented method uses a topology-derived surrogate. Experiments compare placement strategies across communication graphs, aggregation rules, datasets, and data-heterogeneity settings, primarily under Gaussian perturbations.

Critical strengths
  1. The placement problem is relevant to security evaluation. The experiments illustrate that random client selection can substantially understate the vulnerability of some decentralized configurations. Treating attacker placement as an explicit experimental factor is useful independently of the proposed selection method.
  2. The evaluation considers a reasonably broad range of aggregation and topology settings. The comparisons include several robust aggregators, simple centrality baselines, and placement-specific methods. Reporting variability and cases where placement has little effect helps distinguish vulnerabilities associated with particular configurations from a universal failure of decentralized learning.
Critical weaknesses
  1. Theorem 1 does not establish optimality for the stated attack objective. Equation (7) bounds the system-level deviation by a sum over all clients, whereas Appendix E maximizes a sum over the selected clients. These are different quantities. Moreover, nonnegative row-stochasticity implies
\left\|[\Phi(C,T,t)]_{:,k}\right\|_1 =|N|.$$ Consequently, the right-hand side of Eq. (7), with the stated uniform constants, is independent of the compromised set and provides no placement-dependent guarantee. Appendix E only shows that the largest entries maximize a sum of fixed nonnegative scores; it neither transfers optimality from an upper bound to the true deviation nor handles the dependence of the scores themselves on the unknown set $C$. This undermines the central theoretical claim rather than merely requiring a clearer proof presentation. 2. **The practical surrogate is not sufficiently distinguished from conventional centrality.** Equation (10) uses only topology and is identical across aggregation rules and data distributions on the same graph. The implemented algorithm therefore does not instantiate much of the aggregation- and trajectory-dependent information emphasized in the motivation. Many main-table results exactly match degree or PageRank selection. The 1,000-client experiment reports set overlap rather than a corresponding large-scale attack-performance advantage; Appendix F.4.4 states that the proposed method and PageRank select identical sets for every BA seed and four of five ER seeds. Once the unsupported optimality claim is removed, the incremental contribution needs stronger validation.
Questions for authors
  1. Theoretical clarification: Can you provide a set-dependent bound or another argument connecting the selected-client score to the actual objective in Eq. (5), while accounting for the dependence of the mixing matrices on CC? Otherwise, will Theorem 1 be restricted to optimization of an explicitly defined fixed-score surrogate?

  2. Evidence affecting the recommendation: What evidence shows that the topology-only score predicts realized model deviation or attack damage beyond degree/PageRank? In particular, can you provide paired attack-performance comparisons for the larger graphs, rather than only overlap counts, and show where the surrogate remains informative under different aggregators?

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

The paper investigates how compromised-client placement affects attack effectiveness in decentralized federated learning (DFL). It expresses attacked-versus-clean model deviation through products of mixing matrices and derives an upper bound that motivates a client-level accumulated model deviation (AMD) score. Max-CL-AMD selects the mm clients with the highest estimated scores. Because the actual mixing matrices may depend on the compromised set and training dynamics and are generally unavailable at selection time, the implemented method uses a fixed, topology-only approximation. Experiments across multiple communication topologies, aggregation rules, datasets, and IID/non-IID partitions, primarily under Gaussian model perturbations, show that placement can substantially affect attack outcomes. Max-CL-AMD often outperforms random selection, but frequently matches Degree or PageRank and performs worse than random selection in several settings.

Critical strengths
  1. The paper studies compromised-client placement as an important dimension of DFL security. Since malicious updates propagate through graph-dependent, multi-hop paths, separating attacker placement from the poisoning strategy is well motivated. The experiments show that placement can substantially affect attack effectiveness in some settings, highlighting the risk of relying solely on random attacker selection when evaluating robustness.

  2. Max-CL-AMD requires only the communication topology at selection time, avoiding access to training data, model trajectories, or future aggregation weights. The evaluation covers multiple graph families, aggregation rules, datasets, and IID/non-IID partitions, providing useful breadth across experimental conditions. Including both structural baselines, such as Degree and PageRank, and DFL-specific placement methods also helps assess the proposed strategy against relevant alternatives.

Critical weaknesses
  1. The main theorem does not establish optimality. Equation (7)/(28) sums over all clients, while Appendix E introduces a selected-client subset objective without justification. Under the stated nonnegative, row-stochastic assumptions, the original bound is independent of the compromised set 𝐶 . Furthermore, the scores themselves depend on 𝐶 , so fixed-score sorting cannot establish optimal placement. Maximizing an upper bound also does not imply maximizing actual deviation or attack effectiveness. The theorem should be corrected or restricted to a fixed-matrix surrogate.
  2. The algorithm replaces attack- and training-dependent quantities with a fixed topology-only matrix, without establishing approximation accuracy or ranking preservation. No experiments relate estimated scores to actual AMD or TER, isolate matrix-estimation error, or compare against exhaustive optimal placement on small graphs. Moreover, AMD includes compromised clients, whereas TER measures benign-client performance.
  3. Max-CL-AMD frequently matches Degree/PageRank, sometimes underperforms Random, and exhibits substantial variability across only five seeds. Evaluation primarily uses one Gaussian attack family and scale; unconstrained Gaussian perturbations also conflict with the bounded-perturbation assumption. Missing clean TER, incomplete implementation details, and the absence of attack-performance/runtime results for the 1,000-client case further limit the conclusions.
Questions for authors
  1. Can the authors clarify and correct the optimality claim? Under the stated nonnegative, row-stochastic assumptions, Eq. (28) yields a bound independent of C. Can the authors derive a placement-sensitive bound that retains the support of the injected perturbations, or otherwise justify the selected-client objective in Appendix E? Please address the dependence of �0��0� on C and establish its relationship to the actual deviation in Eq. (5). If the result assumes a fixed mixing sequence independent of C, please state this explicitly and restrict the optimality claim to the surrogate property actually proved.

  2. Does the proposed surrogate preserve the ordering of actual attack outcomes? On sufficiently small graphs, please enumerate all feasible size-m sets and compare Max-CL-AMD with the empirically best placements for parameter deviation and TER, using matched randomness and repeated runs. Report estimated optimality gaps and rank correlations across candidate sets between (U^(C)=∑k∈Cω^k(T))(\widehat U(C)=\sum_{k\in C}\widehat\omega_k(T)), the corresponding actual-matrix surrogate where well-defined and observable, realized deviation, and TER. Please distinguish benign-client deviation from all-client deviation. Do these results demonstrate an advantage over Degree, PageRank, and BPI?

  3. How robust and reproducible are the empirical conclusions? Do the findings persist with bounded perturbations consistent with Assumption 3, multiple attack strengths and types, and additional seeds with paired uncertainty estimates? Please provide clean TER, attack-induced �1��1�, per-run outcomes or collapse frequencies, and complete training, topology, aggregation, and tie-breaking settings. For the 1,000-client setting, or a representative larger-scale setting, please report attack effectiveness and selection runtime rather than selection overlap alone. These results would help distinguish reliable gains from baseline ties, occasional collapses, and poor clean-model performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.