acceptodds
Under review as a conference paper at ICLR 2027

Byzantine-Robust Federated Learning across Selection Strategies and Partial-Participation Regimes

Abstract

Federated learning often operates under partial client participation, where only a subset of clients contributes in each communication round. In the presence of Byzantine clients, non-uniform client selection can both bias the observed client population and substantially change the fraction of adversarial clients seen by the server. We propose RaP, a Byzantine-robust federated optimization method for partial participation and general client selection. RaP combines partial-participation bias correction with trial-function-based robust aggregation, validating both fresh updates and the accumulated correction information reused across rounds. The method does not require an honest majority among the clients selected in each round and accommodates stochastic client unavailability. We establish convergence guarantees under heterogeneous client objectives and partial participation. Experiments on heterogeneous CIFAR-10 and real-world ECG classification show that RaP remains stable across uniform and non-uniform client-selection strategies and consistently outperforms Byzantine-robust baselines in challenging adversarial regimes.

Primary area:
optimization
Keywords:
federated learning, byzantine robustness, partial participation, client selection

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

RaP validates both fresh updates and accumulated surrogate gradients using a trusted server objective. It targets Byzantine-robust FL under general client selection and intermittent availability, reporting gains on CIFAR-10 and private ECG data.

Critical strengths

Protecting historical state is a useful observation, supported by Table 3. The CIFAR evaluation covers four selectors and five conditions over five runs, with RaP achieving the highest reported mean accuracy throughout.

Critical weaknesses
  1. Incorrect convergence derivation. On p.34, lines 1807–1810 contain (1−θ)3γδ22(1-\theta)\frac{3\gamma\delta_2}{2}. Dividing by pγ/8p\gamma/8 should produce 12(1−θ)δ2p\frac{12(1-\theta)\delta_2}{p}, but lines 1815–1817 omit it. This nonvanishing heterogeneity residual invalidates the stated derivation of F.5. F.6 also converts the fixed ζ(N)\zeta(N) residual into an additive communication count without the required sample-size/accuracy condition.
  2. Unresolved comparison fairness. Algorithm 2 requires available, non-selected clients to compute surrogates. Communication-adjusted curves do not establish equal computation or final table budgets, and equivalent baseline access to server data/updates is unclear. Under Uniform Random Grad, Fresh-only reaches 0.804 versus BANT's 0.100, leaving most of that gap unexplained by the surrogate. Matched-budget comparisons and baseline tuning details are needed.
  3. Bias correction is not established. Mixing current and historical directions does not itself demonstrate reduced selection bias. The proof bounds the surrogate as a perturbation rather than establishing improved alignment or convergence. Its sufficient conditions restrict CIFAR's θ\theta to approximately 3.3×10−73.3\times10^{-7} or less, versus 0.15 in experiments. Table 3 supports usefulness, but does not isolate bias correction from extra information or scaling effects.
  4. Insufficient targeted robustness tests. Main comparisons use 20% Byzantine clients without reporting malicious-majority cohort frequencies. No surrogate-only attack with honest fresh updates or attack explicitly adapted to trial-function validation is reported. The CIFAR availability sweep lacks matching reduced-availability baselines; ECG baseline missing-response handling is unspecified.
  5. Unclear technical increment. The differences from BANT and Bylinkin et al. (2025) need explicit algorithmic and analytical comparison. Applying conventional robust aggregators to the same surrogates is a missing control for whether trial-function validation is essential. Honest-majority-free aggregation alone does not distinguish RaP from the cited trial-function literature.
Questions for authors

Please refer to weakness.

Reviewer_ojddOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper considers the problem of Byzantine-robustness in FL under partial client participation, wherein only some non-uniformly selected clients participate in each synchronization round. To solve the problem the paper proposes trust-function based robust aggregation. The convergence of the algorithm is formally analyzed, and presented empirical evidence shows improvement over some existing schemes.

Critical strengths
  1. Problem is timely and important. The problem of robustness under partial client participation is important for practical purposes.
  2. Extensive empirical evaluation. The paper presents a decent empirical evaluation of the proposed robustness scheme.
Critical weaknesses
  1. The algorithm description can be improved. The quantities like smks_m^k, the arg min over w∈Δ1Mw \in \Delta_1^M in Lines 11 and 15, the role of θ\theta are introduced without explanation, and the client-side procedure is deferred entirely to the appendix.
  2. The improvement over existing trial-function-based methods is not established. The abstract claims that RaP combines partial-participation bias correction with trial-function-based robust aggregation, but it is not shown how this improves over the trial-function approach it builds on (e.g. BANT). The accompanying claim of convergence guarantees under heterogeneous objectives and partial participation reads as decorative rather than as a substantive contribution in the current form.
  3. The theoretical contribution is weak. Assumption 3.3 imposes a bounded-heterogeneity relation between the trusted population objective f1f_1 and ff, which raises the question of why one would not simply minimize f1(x)f_1(x) directly. In Theorem 3.5, the quantity Δ\Delta is never explained and the bound exhibits no visible dependence on the Byzantine fraction. If ζ(N)∈O(1/N)\zeta(N) \in \mathcal{O}(1/N), the stated rate can be obtained simply by minimizing over f1(x)f_1(x), which diminishes the need for federation in the first place.
  4. The threat model is incomplete. (i) Byzantine clients can choose their updates vmv_m so as to minimize the trial function f^\hat f and thereby receive large weights in Line 15. (ii) The mean availability qm∈(0,1]q_m \in (0,1] is treated as given, but a Byzantine client could simply report qm=1q_m = 1.
  5. The treatment of non-selected clients is unclear. Non-selected clients may still receive the broadcast and maintain their local surrogate state (lines 150–152); if so, why don’t they return their updates then?
  6. Relevant baselines are missing. The comparison does not include nearest-neighbor mixing (NNM) [a] or adaptive robust clipping (ARC) [b], both standard enhancements for distance-based robust aggregation.

References:

[a] Allouah, Youssef, et al. "Fixing by mixing: A recipe for optimal byzantine ml under heterogeneity." International Conference on Artificial Intelligence and Statistics. PMLR, 2023.

[b] Allouah, Youssef, et al. "Adaptive gradient clipping for robust federated learning." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors

Please respond to the above weaknesses and questions therein.

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

The paper proposes RAP for Byzantine-robust federated learning with partial participation and general client selection. It uses a trusted server-side objective to validate both fresh updates and accumulated surrogate corrections. Experiments evaluate the method across several selection strategies and attacks on CIFAR-10 and ECG classification.

Critical strengths

The paper identifies a useful problem: corrupted historical corrections can undermine training even when fresh updates are protected. Table 3 directly supports the need to validate both components. The evaluation covers several client selectors, and the ECG experiment uses an external public dataset for server-side validation.

Critical weaknesses

The convergence guarantee is confusing and does not look like having a rigorous proof.

The experiments are conducted on the small-scale CIFAR-10 dataset and ECG dataset with simple attacks, which does not look interesting.

Questions for authors

Is there an additional assumption that excludes this example? Otherwise, how should Theorem F.5 and Theorem 3.5 be revised to account for the heterogeneity floor? Please also clarify the attainable accuracy for a fixed trusted-set size.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.