- 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
-
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.
-
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
-
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.
-
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
-
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?
-
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.
-
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.
-
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.