acceptodds
Under review as a conference paper at ICLR 2027

FedR²DG: Reconciling Adversarial Robustness and Domain Generalization with Selective Defense in Federated Learning

Abstract

Federated adversarial training becomes difficult when clients differ in both data domain and defense budget, especially when the deployment domain is unseen during training. We study selective federated robust domain generalization, where only a subset of source clients can afford full adversarial training and no target-domain data are available during training. We derive a held-out robust-risk bound that separates source-local fit, a federated mechanism gap, class-conditional source coverage, label-prior mismatch, and a conditional residual. Motivated by this bound, we propose Federated Reconciliation of Adversarial Robustness and Domain Generalization (FedR²DG), which combines budget-aware local defense with drift control and class-conditional alignment. Experiments on PACS, OfficeHome, and mini-DomainNet50 show that FedR²DG improves aggregate robust accuracy against AutoAttack by 4.91–11.99 percentage points over the best baseline in each dataset–budget setting. It also achieves the highest harmonic mean of standard and robust accuracy (H-score) among the compared methods, while largely preserving standard accuracy. The full-run computational cost per weakly defended client is 1.22–1.43× that of the matched selective-defense baseline. These results indicate that FedR²DG improves the standard-robust balance under heterogeneous defense budgets without requiring full adversarial training at every client.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
domain generalization, federated learning, adversarial robustness, distribution shift

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_djSoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper considers cross-institution federated learning, where source clients belong to distinct data domains and only a subset of institutions can afford full adversarial training (AT), while the deployment domain is unseen during training. Spanning the three subfields of federated optimization, adversarial training, and domain generalization, the authors formalize this setting as Selective Federated Robust Domain Generalization (SFRDG) and propose FedR²DG, which trains defended clients with PGD, handles weakly defended clients via replay-based approximate adversarial training, and further incorporates global class-centroid alignment and aggregation-weight rebalancing.

Critical strengths

1.The paper is well-organized, with a sound motivation and clear presentation. 2.The paper provides comprehensive details for reproducibility. 3.The theoretical analysis is rigorous and the derivation chain is complete; the correspondence between the theoretical components and the proposed method is clearly traceable, which lends credibility to the method's design rationale

Critical weaknesses

1.The paper claims federated robust domain generalization, yet contributes no new domain-generalization mechanism and compares against no domain-generalization method, so it is impossible to see what the paper advances on the domain-generalization axis. 2.The two client types share the same perturbation radius ε and differ only in attack quality and compute, so the paper proposes no new adversarial training method. 3.The paper states no threat model and specifies no AutoAttack configuration (version, attack list, steps, restarts), which is an evaluation-standards deficiency: AA results are version- and step-sensitive, so the reported robustness numbers can be neither reproduced nor judged sufficient. 4.The baselines are all generic federated optimization methods, with no comparison against the closest recent work — partial adversarial training (SFAT/SFAT*) or single-AT-client FAT (FAML) — nor against an all-clients-AT upper bound. 5.The paper does state in §5.1 that baseline weak clients train on clean data only, but it never treats this as a confound or provides an equal-budget replay control (Table 2 shows both cost the same 7.89 TMACs per round), so the source of the performance gains cannot be attributed. 6.The paper contains no method/architecture diagram and no performance or training-dynamics plots, leaving key claims without visual support.

Questions for authors

1.Since the method introduces no new domain-generalization mechanism and no domain-generalization method is used for comparison, could you clarify what the paper advances specifically on the domain-generalization axis? 2.Apart from the standard L-infinity inner maximization and PGD/Free-AT-style training shared by all compared methods, does FedR²DG introduce any new attack or defense mechanism? 3.What is the exact threat model, and what AutoAttack configuration was used — which version, which attacks, how many steps and restarts — given that neither is specified anywhere in the paper? 4. Could you add comparisons against the closest recent work — partial adversarial training (SFAT/SFAT*) and single-AT-client FAT (FAML) — together with an all-clients-AT upper bound? 5.Since FedR²DG's weak clients perform replay while all baselines' weak clients train on clean data only, at identical cost (7.89 TMACs/round, Table 2), could you provide an equal-budget replay control and report the weak clients' local robust accuracy?

6 comments
AuthorsOpenReview ↗

Response to Reviewer djSo (1/2)

We thank the reviewer for the positive comments on clarity and reproducibility.

W1 / Q1. Contribution on the DG axis

Our DG-side contribution is a robust-risk analysis for source-only robust DG and the class-conditional alignment it motivates: the bound conditions the robust loss on class-conditional clean features, so we align clean source features of each class around shared anchors. We now compare against FedSR, FedDG-GA, and FedOMG under the same selective defense (Table R1). These methods improve clean generalization but lower mean held-out robust accuracy below SR-FedAvg, so better clean DG does not by itself carry the defended client's robustness to the unseen domain. In DG, where the target is unseen, avoiding collapse in any case matters: FedR²DG achieves the highest H-score in every defended-client case, while every baseline falls to at most 18.6% AA when Sketch is the defended client.

W2 / Q2. New attack or defense mechanism

FedR²DG does not propose a new inner maximization; all clients share the same ℓ∞ threat model and ε. Our defense-side contribution is how budget-limited clients take part in robust training. We adapt Free-AT (Shafahi et al., 2019), which obtains parameter and input gradients from one backward pass, to the federated setting: perturbations reset for each mini-batch and round instead of being carried across mini-batches as in Free-AT, because a perturbation crafted for other images under an earlier model is not a meaningful starting point, so each replay sequence is a short sign-gradient ascent from zero on its own images under the current model; only part of each round's mini-batches are replayed, and every update is followed by EMA tethering. Weakly defended clients thus perform proxy adversarial training at the per-round cost of clean training. To our knowledge, this is the first use of proxy adversarial training at the per-round cost of clean training for weakly defended clients when robustness must transfer to an unseen domain. We then address the clean degradation it causes through prototype alignment (Table R2).

The adaptation itself matters, not only the amount of adversarial training: with equal or less compute, replay alone (C1: 43.83% AA, 56.60 H) outperforms spending the budget on PGD-10 updates (38.06 / 49.44) or on 10 clean steps plus one PGD-10 step (42.37 / 55.79; see W5/Q5 below), and it prevents the collapse with Sketch as the defended client (37.60% vs. 2.46% AA for SR-FedAvg).

Existing partial-AT aggregation does not resolve this setting. SFAT* reweights only the adversarially trained clients according to their adversarial training loss, so it cannot change how much robustness the weakly defended domains contribute; it collapses when Sketch is the defended client (2.9% AA in Table R1; 2.5–6.8% with two or more defended clients), and SR-FRP, which propagates robustness through BN statistics, reaches 6.6% AA. We attribute this failure to domain heterogeneity itself: robustness learned on one domain does not carry over to the others. Proxy adversarial training addresses this at its source by letting every weakly defended client fit robustness on its own domain, which raises their local models' AA from 11.0% to 36.4%. The margin over these methods (45.6% vs. 6.6–28.6% AA) is far larger than the differences among our components.

W3 / Q3. Threat model and AutoAttack configuration

We consider white-box ℓ∞ evasion attacks at test time on the held-out domain: the attacker has full access to the final global model, including its gradients, and the radius equals the training radius ε of each setting (2/255, 4/255, or 8/255; 8/255 in all new experiments unless noted otherwise).

Robust accuracy, denoted AA in the paper, is computed with the official AutoAttack implementation (Croce & Hein, ICML 2020; fra31/auto-attack v0.1) in its custom mode with its two gradient-based attacks, APGD-CE and APGD-T; FAB-T and Square are not included. All attack parameters are the library defaults (n_iter = 100 for both attacks, n_restarts = 5 for APGD-CE and 1 for APGD-T, eot_iter = 1, ρ = 0.75), and we fix seed = 42 for reproducibility. The only dataset-dependent setting is the number of target classes for APGD-T, min(9, C − 1): 6 for PACS and 9 for OfficeHome and mini-DomainNet50. The same protocol is applied to FedR²DG and to every baseline.

In the revision, we will state the threat model and this configuration in Section 5.2 and make clear in the text and table captions that AA denotes APGD-CE + APGD-T rather than the full standard AutoAttack ensemble.

AuthorsOpenReview ↗

Response to Reviewer djSo (2/2)

W4 / Q4. SFAT/SFAT*, FAML, and an all-clients-AT upper bound

We add SFAT*, the partial federated AT extension of SFAT (Zhu et al., TPAMI 2026), using its official implementation (Table R1). The original SFAT (Zhu et al., ICLR 2023) assumes that every client performs AT, whereas SFAT* aggregates standard-training and adversarially trained clients separately and then combines them, which matches our selective-defense setting. In the official implementation, the α-slack reweights only the adversarially trained clients, so with one defended client it is inactive and SFAT* reduces to uniform averaging; it reaches 28.6% AA versus 45.6% for FedR²DG. In the configurations with two or more defended clients, where its slack is active (6/2, 9/3, and 9/2 in Table R4 of the General Response), SR-SFAT* reaches 27.0–30.6% AA versus 46.0–46.6% for FedR²DG. FedAvg with PGD-10 AT on every client reaches 49.88% AA and 61.73 H; FedR²DG recovers about 88% of the H-score gain while each weakly defended client spends about 1/11 of the per-round cost of PGD-10 training (Table R3).

FAML (Hossain et al., CVPR 2026 Findings) combines federated adversarial training with meta-learning under data scarcity and hardware heterogeneity, and reports that adversarial training on a single client can suffice, which is closely related to selective defense. Its clients partition a single dataset, and its out-of-domain results evaluate a model trained on one dataset on another, whereas SFRDG requires transferring robustness from domain-heterogeneous source clients to a held-out domain. Because its code is not publicly available (the repository linked from its project page was empty as of November 13, 2026), we could not reproduce it faithfully and did not include it as an empirical baseline; we will add it to the related work and discuss it in the revision.

W5 / Q5. Equal-budget replay control and weak clients' local robust accuracy

Replay is part of our method by design rather than a confound: it lets weakly defended clients contribute robustness at the same per-round cost as clean training (7.89 TMACs in both methods, Table 2). To attribute the gains, C1 in Table R2 is SR-FedAvg with the identical replay schedule and all other components disabled: replay alone raises AA from 27.17% to 43.83% (C0 → C1), and the remaining components add +3.4 pp clean accuracy, +1.8 pp AA, and +2.3 H on top of it (C1 → C4); Table R2 further shows the contribution of each component. We also give the baselines' weakly defended clients the same per-round budget for PGD-10 training:

\hline \text{Weakly defended client (per-round budget of 20 clean updates)} & \text{Clean} & \text{AA} & \text{H} \\\\ \hline \text{SR-FedAvg, 2 PGD-10 updates} & 79.92 & 38.06 & 49.44 \\\\ \text{SR-FedAvg, 10 clean updates + 1 PGD-10 update} & 82.97 & 42.37 & 55.79 \\\\ \text{SR-FedAvg + replay (C1)} & 80.58 & 43.83 & 56.60 \\\\ \text{FedR}^2\text{DG}\text{ (C4)} & 83.97 & 45.59 & 58.89 \\\\ \hline \end{array}$$ One PGD-10 update costs about 11 clean updates (Table 2: 86.79 vs. 7.89 TMACs per 20 updates), so both PGD allocations use slightly more compute (1.05–1.1×) than replay. With equal or less compute, replay alone (C1) already outperforms both PGD allocations, and FedR²DG is the strongest. Clean and robust accuracy of each weakly defended client's final local model on the held-out target (mean over six client-case pairs: two weakly defended clients × three defended-client choices): $$\begin{array}{lcc} \hline \text{Weakly defended clients' local models} & \text{Clean} & \text{AA} \\\\ \hline \text{SR-FedAvg (clean CE) (C0)} & 83.80 & 11.03 \\\\ \text{SR-FedAvg + replay (C1)} & 80.10 & 36.42 \\\\ \text{FedR}^2\text{DG}\text{ (C4)} & 81.34 & 37.16 \\\\ \hline \end{array}$$ Replay makes weakly defended clients contribute robust updates rather than only clean ones. Because local models start each round from the broadcast global model, part of their robustness is inherited from it, so we use this comparison across methods rather than as an absolute measure. **W6. Figures** Figure 1 illustrates the problem setting, and Figure 2 reports standard and robust accuracy across hyperparameters. We will additionally add a method overview figure and further visualizations of the results.
AuthorsOpenReview ↗

Follow-up for Reviewer djSo

Dear Reviewer djSo,

Thank you again for your review. We have uploaded a revised PDF (changes are marked in blue and summarized in General Response 2/2), which adds the threat model and the exact AutoAttack configuration (Section 5.2, Appendix F.3.1), SFAT*, FAML, and federated DG methods to the related work, the all-clients-AT upper bound (Table 2), the equal-budget comparison (Appendix E.3), a method overview figure (Figure 1), and client drift over training rounds (Figure 3). Could you let us know whether any of your concerns remain unresolved? We would be glad to address them before the discussion period ends.

Reviewer_g5rcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a practical and important problem of selective federated robust domain generalization (SFRDG), where source clients differ in their data domains and defense budgets, and the global model should generalize robustly to an unseen target domain. The authors propose FedR^2DG, a well-motivated framework that combines budget-aware local defense, parameter-drift control, class-conditional prototype alignment, and aggregation reweighting. A robust-risk bound decomposes target risk into source-local robust risk, federated drift, class-conditional coverage, label-prior mismatch, and a conditional residual term. Experiments on PACS, OfficeHome, and mini-DomainNet50 evaluate clean and robust accuracy under three perturbation budgets. FedR^2DG improves aggregate robust accuracy and the standard–robust harmonic mean over the evaluated baselines.

Critical strengths

This paper addresses an important and practical problem: adversarial robustness and domain generalization in selective federated learning.

This theoretical formulation is clear in sec. 3 and the framework is well-motivated .

The experimental validation in this paper is comprehensive.

Critical weaknesses

The paper does not examine or discuss the robustness and computational cost disparity between the default setting and the setting where all clients are defended clients (not the weakly defended clients) in sec. 5. And the defense ratio of clients is fixed, limiting claims about selective defense.

Due to the lack of component ablation in sec. 5, the roles of drift control, class-conditional alignment, and aggregation reweighting are unclear.

Compared with the baselines using clean CE in weakly defended clients, the weakly defended clients in FedR^2DG seem to include extra replay at a significantly computational cost. Do the results reported in tab. 2 also include replay for SR-FedAvg? And the warmup setting for baselines is unclear.

Questions for authors

How much of the improvement remains when SR-FedAvg receives the identical replay schedule, with all other FedR^2DG components disabled? Can existing results isolate this comparison?

Are the warmup and weakly defended settings for the baselines the same as those for FedR^2DG?

Does the practical computational budget in tab. 2 account for periodic centroid updates, communication, and synchronization, or only local training?

5 comments
AuthorsOpenReview ↗

Response to Reviewer g5rc (1/2)

We thank the reviewer for the positive assessment of the problem and formulation.

W1. All clients defended; fixed defense ratio

With PGD-10 AT on every client, FedAvg reaches 80.96 / 49.88 / 61.73 (Clean / AA / H), versus 83.97 / 45.59 / 58.89 for FedR²DG (Table R3 in the General Response). FedR²DG thus recovers about 88% of the H-score gain from SR-FedAvg to this upper bound, while each weakly defended client spends 7.89 instead of 86.79 TMACs per round (about 1/11). Selective defense trades 4.3 pp of AA for this cost reduction while keeping 3.0 pp higher clean accuracy.

For different defense ratios (1/9 to 1/3, with 6 or 9 source clients; Table R4), FedR²DG achieves the highest AA and H-score in every configuration, exceeding the second-best method by 15.4–19.5 pp AA. Its AA stays at 45.7–46.9%, whereas SR-FedAvg drops from 27.6–28.1% at a ratio of 1/3 to 19.1% at 1/9.

W2. Component ablation

Please see Table R2 in the General Response. Replay provides robustness, warm-up and prototype alignment recover clean accuracy, and √n weighting (over data-proportional weighting) and EMA tethering each improve the H-score. For drift control, EMA tethering at the default γema\gamma_{\mathrm{ema}} = 0.1 reduces the mean parameter drift of clients from the broadcast global model by about 50% (0.276 → 0.129) compared with no tethering (see the γema\gamma_{\mathrm{ema}} table in our response to Reviewer Bvvy, Q1).

W3 / Q1. Replay cost and identical replay schedule for SR-FedAvg

Replay is part of our method by design rather than a confound: it lets weakly defended clients contribute robustness at the same per-round cost as clean training (7.89 TMACs in both methods, Table 2). Table 2 reports SR-FedAvg without replay; the per-round costs are identical because each replay update uses one forward and one backward pass, as a clean step does. Over the full run, a weakly defended client's cost is 1.22–1.43× that of SR-FedAvg (Table 2); this difference comes from centroid computation, not from replay.

Our design also differs from partial federated AT such as SFAT*, which keeps adversarially trained and standard-training clients as separate groups and aggregates them hierarchically. In FedR²DG, every client takes part in robust training within its own budget, so all clients are aggregated by a single weighted average (Eq. (12)) without being separated by defense type. Involving all clients in robust training through this unified framework is therefore the intended design and part of our contribution (General Response, Positioning (ii)).

Existing results isolate this comparison: we evaluate SR-FedAvg with the identical replay schedule and all other components disabled (C1 in Table R2). Replay alone raises AA from 27.17% to 43.83%, and the remaining components add +3.4 pp clean accuracy, +1.8 pp AA, and +2.3 H on top of it (C1 → C4). We also give the baselines' weakly defended clients the same per-round budget for PGD-10 training:

\hline \text{Weakly defended client (per-round budget of 20 clean updates)} & \text{Clean} & \text{AA} & \text{H} \\\\ \hline \text{SR-FedAvg, 2 PGD-10 updates} & 79.92 & 38.06 & 49.44 \\\\ \text{SR-FedAvg, 10 clean updates + 1 PGD-10 update} & 82.97 & 42.37 & 55.79 \\\\ \text{SR-FedAvg + replay (C1)} & 80.58 & 43.83 & 56.60 \\\\ \text{FedR}^2\text{DG}\text{ (C4)} & 83.97 & 45.59 & 58.89 \\\\ \hline \end{array}$$ One PGD-10 update costs about 11 clean updates (Table 2: 86.79 vs. 7.89 TMACs per 20 updates), so both PGD allocations use slightly more compute (1.05–1.1×) than replay. With equal or less compute, replay alone (C1) already outperforms both PGD allocations, and FedR²DG is the strongest. **Q2. Warm-up and weakly defended settings of baselines** Weakly defended clients in all baselines use clean CE with 20 local steps per round, as in FedR²DG's step budget. Baselines do not use warm-up because its only role in FedR²DG is to initialize class centroids for the prototype loss. All methods use the same 250 rounds and 5,000 local steps per client, so warm-up replaces 50 rounds of selective-defense training with clean training rather than adding training, and the defended client's total cost is lower (0.91–0.93×, Table 2). To isolate warm-up itself, we compare FedR²DG without the prototype loss with and without warm-up: $$\begin{array}{lccc} \hline \text{Setting (}\lambda\text{=0)} & \text{Clean} & \text{AA} & \text{H} \\\\ \hline \text{Without warm-up} & 80.04 & 46.05 & 58.27 \\\\ \text{With warm-up} & 82.53 & 45.75 & 58.68 \\\\ \hline \end{array}$$ Warm-up mainly raises clean accuracy (+2.5 pp) and changes AA by −0.3 pp; it does not explain the robustness gap to the baselines.
AuthorsOpenReview ↗

Response to Reviewer g5rc (2/2)

Q3. Scope of the cost in Table 2

Table 2 counts local computation (in TMACs) only: local training, including warm-up, and the centroid computation at initialization and at all 40 refreshes. Communication and synchronization are not included, because the constraint that SFRDG targets is each client's computational budget rather than its bandwidth or latency.

For reference, the extra communication of FedR²DG is small: each centroid refresh adds C × 2048 values per client for the uploaded local centroids and the same amount for the broadcast global centroids (7 × 2048 ≈ 14K for PACS, 65 × 2048 ≈ 133K for OfficeHome, 50 × 2048 ≈ 102K for mini-DomainNet50), about 0.06%, 0.6%, and 0.4% of the 23.52M model parameters that every method exchanges each round, and refreshes occur only every 5 rounds. Combining these vectors on the server costs O(K × C × 2048) operations for K clients, negligible next to averaging the model parameters. We will state this scope explicitly in Section 5.4.

AuthorsOpenReview ↗

Follow-up for Reviewer g5rc

Dear Reviewer g5rc,

Thank you again for your review. We have uploaded a revised PDF (changes are marked in blue and summarized in General Response 2/2), which adds the component ablation and the all-clients-AT upper bound (Section 5.3, Table 2), the equal-budget comparison (Appendix E.3), results with lower defense ratios (Appendix E.4), and the scope of the cost analysis (Section 5.4). Could you let us know whether any of your concerns remain unresolved? We would be glad to address them before the discussion period ends.

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

This paper proposes FedR$^2$DG for federated adversarial training with heterogeneous defense budgets across clients and generalization to unseen target domains. Firstly, the authors derive a held-out robust-risk bound that decomposes the target-domain risk into source robust fit, aggregation drift, class-conditional feature discrepancy, label-prior mismatch, and a conditional residual, together with finite-sample corrections. Motivated by this bound, the proposed method combines budget-aware local adversarial training, drift control, and class-conditional prototype alignment. Experiments on several multi-domain benchmarks show improved adversarial robustness over the evaluated baselines.

Critical strengths
  1. The paper provides a detailed decomposition of the held-out robust-risk bound, clarifying why improving adversarial robustness on source clients alone is insufficient to guarantee robust generalization to unseen target domains. Meanwhile, this bound helps to design a reliable method.
  2. The paper clearly motivates the joint challenges of heterogeneous defense budgets across clients and robust generalization to unseen domains. This setting is practically relevant, as clients may have different computational resources and deployment distributions may differ from those observed during training.
Critical weaknesses
  1. This paper focuses on combined problems, different client defense budgets and robust generalization to unseen domains, which existing work have studied respectively. These directions have been explored separately: FRP [1] studies robustness propagation from resource-rich to resource-limited clients, while Zou and Liu [2] investigate adversarial robustness under out-of-distribution generalization. Federated generalization to unseen domains has also been studied by FedSR [3]. Although the combined setting is relevant, the paper does not sufficiently clarify the additional technical challenges when encountering a new setting.
  2. Although the paper decomposes the target-domain adversarial risk bound and motivates its components accordingly, it does not provide sufficient empirical evidence showing how these components affect the corresponding terms in the bound. Moreover, the ablation study does not isolate the robustness gains attributable to each component, particularly parameter tethering in Eq.(8), weak-client replay in Eq.(11), and aggregation weighting in Eq.(12).

[1] Junyuan Hong, Haotao Wang, Zhangyang Wang, and Jiayu Zhou. Federated Robustness Propagation: Sharing Adversarial Robustness in Heterogeneous Federated Learning. AAAI, 2023.

[2] Xin Zou and Weiwei Liu. On the Adversarial Robustness of Out-of-distribution Generalization Models. NeurIPS, 2023.

[3] A. Tuan Nguyen, Philip Torr, and Ser Nam Lim. FedSR: A Simple and Effective Domain Generalization Method for Federated Learning. NeurIPS, 2022.

Questions for authors
  1. γema\gamma_{ema} in Eq.(8) controls how strongly local parameters are pulled toward the broadcast global model at each update. However, authors did not report a sensitivity analysis or an ablation with γema=0\gamma_{ema} = 0. Additionally, it would be better to report the extent to which the client models deviates global models during training to verify the effectiveness of eq.(8).
  2. The weight define in Eq.(12) depend only on client sample sizes and do not account for different defense budgets (or robustness of client models). It remains unclear how the proposed method handle the resulting imbalance in client robustness. Additionally, authors should compare different weighting schemes to explain why eq.(12) is chosen.
  3. The prototype loss directly aligns clean representations, whereas the central challenge is transferring adversarial robustness to unseen domains. In the experimental results about hyperparameter λ\lambda, removing the prototype loss (λ=0\lambda = 0) appears to have little effect on robust accuracy, while larger λ\lambda eventually degrade it. Eq.(3) theoretically states that prototype alignment helps to decrease the robust risk of target domain, but empirical evidence does not give consistent conclusion. The clearest benefit from experimental evidence is instead an improvement in clean accuracy and the current evidence does not establish that prototype alignment contributes to robustness transfer itself. It seems that reducing the term about prototype alignment does not guarantee a reduction in target-domain robust risk, as other terms may also change. Please authors further clarify the function of prototype alignment and why robust accuracy reduces involving prototype alignment.
  4. The authors compare their method with only four baselines, none of which specifically targets unseen-domain generalization. Please include comparisons with federated domain generalization methods such as FedSR [1], FedDG-GA [2], and FedOMG [3].

[1] A. Tuan Nguyen, Philip Torr, and Ser Nam Lim. FedSR: A Simple and Effective Domain Generalization Method for Federated Learning. NeurIPS, 2022. [2] Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Federated Domain Generalization with Generalization Adjustment. CVPR, 2023. [3] Binh Nguyen, Minh-Duong Nguyen, Jinsun Park, Viet Pham, Won-Joo Hwang. Federated Domain Generalization with Data-free On-server Matching Gradient. ICLR, 2025.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Bvvy (1/3)

We thank the reviewer for the detailed questions on the link between the bound and the method.

W1. Additional challenges of the combined setting

FRP studies robustness propagation among participating clients, Zou & Liu study the robustness of centralized OOD models, and FedSR studies clean federated DG. Combining these settings introduces two difficulties that none of them addresses. First, robustness from a few defended clients transfers unevenly to the unseen domain: SR-FedAvg reaches 45.3% AA when Art is the defended client but 2.5% when Sketch is. Second, applying methods from each line does not resolve this: with the same selective adversarial training, FL+DG methods including FedSR improve clean accuracy but lower mean held-out robust accuracy below SR-FedAvg, and SR-FRP, whose BN-based propagation cannot use target-domain statistics in this setting, falls to 6.6% AA (Table R1). FedR²DG addresses both: replay lets every weakly defended client fit robustness on its own domain, and prototype alignment preserves the clean class structure. It achieves the highest H-score in every defended-client case.

W2. Empirical evidence linking components to bound terms

Please see Table R2 in the General Response, which adds the components one at a time and thus isolates the robustness gain of each: replay raises AA by 16.7 pp, and √n weighting and EMA tethering raise it by 1.7 and 1.6 pp. The bound terms targeted by EMA tethering and √n weighting are reported in our answers below: client drift (Q1) and the label-prior radius r^Y\hat{r}_Y (Q2).

Q1. Sensitivity to γema\gamma_{\mathrm{ema}} and client drift

\hline \gamma_{\mathrm{ema}} & 0 & 0.05 & 0.1 & 0.2 & 0.3 \\\\ \hline \text{Clean} & 85.31 & 84.55 & 83.97 & 82.81 & 82.14 \\\\ \text{AA} & 43.97 & 44.57 & \mathbf{45.59} & 45.45 & 44.23 \\\\ \text{H} & 57.78 & 58.17 & \mathbf{58.89} & 58.53 & 57.33 \\\\ \text{Mean parameter drift} & 0.276 & 0.182 & 0.129 & 0.077 & 0.053 \\\\ \text{Mean BN-buffer drift} & 16.61 & 16.61 & 16.54 & 16.42 & 16.34 \\\\ \hline \end{array}$$ Drift is the ℓ2 distance between each client's model at the end of local training in a round (before aggregation) and the global model broadcast at the start of that round, computed separately over all trainable parameters and over the BN running statistics, and averaged over the 200 rounds after warm-up, all clients, and the three defended-client choices. Without tethering ($\gamma_{\mathrm{ema}}$ = 0), robustness is lower (AA 43.97% vs. 45.59%, H 57.78 vs. 58.89 at the default $\gamma_{\mathrm{ema}}$ = 0.1), and tethering at $\gamma_{\mathrm{ema}}$ = 0.1 reduces the mean parameter drift by about 50% (0.276 → 0.129). This decrease holds in each defended-client case and for both the defended and the weakly defended clients. BN-buffer drift, which Eq. (8) does not tether, changes by less than 2% across $\gamma_{\mathrm{ema}}$ (16.61 → 16.34). We will add this analysis to the revision.
AuthorsOpenReview ↗

Response to Reviewer Bvvy (2/3)

Q2. Weighting and robustness imbalance

Eq. (12) does not target robustness imbalance, which we handle locally: weakly defended clients perform replay at the per-round cost of clean training (their local models' target AA rises from 11.0% to 36.4%), and EMA tethering limits each client's drift. Weighting by client robustness would require adversarial evaluation at weakly defended clients, which conflicts with their budget. SFAT*, which adjusts aggregation weights among adversarially trained clients through its α-slack mechanism, reaches 27.0–30.6% AA versus 46.0–46.6% for FedR²DG in the configurations where its slack is active (two or three defended clients; Table R4 in the General Response).

Which source mixture best matches an unseen target cannot be determined without target data. The aggregation weights ν appear in several terms of the bound. Among the terms tied to the source mixture, the label-prior radius r^Y(ν)\hat{r}_Y(\nu) in Theorem 1 depends only on source class counts, whereas the mixture-weight mismatch ∣αc−νc∣1\\|\alpha_c-\nu_c\\|_1 in Corollary 1, where αc\alpha_c are the weights of a source mixture that approximates the target class-conditional distribution, cannot be estimated without target data (Remark 2).

We set the weights using only source information and adopt the geometric mean of data-proportional weighting (every sample counts equally, as in FedAvg) and uniform weighting (every client counts equally) as a compromise between the two, rather than an optimized choice. Label-prior radius r^Y\hat{r}_Y, computable from source data alone:

\hline \text{Held-out target} & \text{Data-proportional} & \sqrt{n}\text{ (ours)} & \text{Uniform} \\\\ \hline \text{PACS A} & 0.4385 & 0.3846 & 0.3999 \\\\ \text{PACS C} & 0.4091 & 0.3844 & 0.4450 \\\\ \text{PACS P} & 0.3235 & 0.3202 & 0.3578 \\\\ \text{PACS S} & 0.1998 & 0.1901 & 0.1896 \\\\ \text{OfficeHome A} & 0.2142 & 0.2140 & 0.2138 \\\\ \text{OfficeHome C} & 0.3218 & 0.2991 & 0.2732 \\\\ \text{OfficeHome P} & 0.2963 & 0.2759 & 0.2530 \\\\ \text{OfficeHome R} & 0.3495 & 0.3250 & 0.2971 \\\\ \hline \end{array}$$ Each extreme yields the largest $\hat{r}_Y$ for some targets (data-proportional for six, uniform for two), whereas √n weighting never does; across the eight targets, √n gives the smallest worst-case $\hat{r}_Y$ (0.385 vs. 0.439 for data-proportional and 0.445 for uniform) and the smallest mean $\hat{r}_Y$ (0.299 vs. 0.319 and 0.304). In accuracy on PACS-P, √n improves over data-proportional weighting (with all other components of FedR²DG fixed: AA 43.31 → 45.59, H 56.66 → 58.89), and uniform weighting performs best (AA 47.37, H 60.72). We attribute this to source-target similarity rather than the label prior: compared with √n, uniform weighting gives more weight to Art and Cartoon, the sources that transfer best to Photo (standard accuracy of single-source Robust ERM on Photo, Table 13: 80.2% and 72.0% vs. 45.4% for Sketch), and less weight to Sketch. Such similarity is unknown in DG, so which extreme wins can only be determined after the fact and can change with the target. On PACS-C and PACS-S, by contrast, √n is slightly better than uniform weighting (AA 47.10 vs. 46.86 and 69.77 vs. 68.85; H 58.74 vs. 58.29 and 75.68 vs. 74.96). Because the target domain is unseen, the best weighting cannot be known in advance, and we believe that a DG method should therefore favor stability across targets over the best choice for any single target. √n weighting follows this principle by avoiding the worst case of either extreme, and we do not claim that it is better for every target. On mini-DomainNet50, all source domains have equal size and the three schemes coincide.
AuthorsOpenReview ↗

Response to Reviewer Bvvy (3/3)

Q3. Role of prototype alignment

Eq. (3) is an upper bound; it does not imply that lowering LprotoL_{\mathrm{proto}} alone must reduce target robust risk, since other terms may change jointly (Remark 1). In Theorem 1, Z is the normalized feature of the clean input, and the robust-loss profile is conditioned on this clean feature; the prototype term therefore arises on clean features, and we apply it to clean features accordingly. Its role is to preserve the clean class-conditional structure under adversarial training.

Preserving clean accuracy is aligned with, not separate from, the robust objective. Following Zhang et al. (2019), for the 0–1 loss, which AA measures, the robust risk of a classifier hh at radius ε\varepsilon decomposes exactly as

Rrob(h)=Pr⁡[h(X)≠Y]⏟natural (clean) risk Rnat(h)+Pr⁡[h(X)=Y,∃,∣δ∣∞≤ε:h(X+δ)≠Y]⏟boundary risk Rbdy(h).\mathcal{R}_{\mathrm{rob}}(h)=\underbrace{\Pr\big[h(X)\neq Y\big]}_{\text{natural (clean) risk } \mathcal{R}_{\mathrm{nat}}(h)}+\underbrace{\Pr\big[h(X)=Y,\\ \exists\\,\\|\delta\\|_\infty\le\varepsilon:\\ h(X+\delta)\neq Y\big]}_{\text{boundary risk } \mathcal{R}_{\mathrm{bdy}}(h)}.

Since Rbdy(h)≥0\mathcal{R}_{\mathrm{bdy}}(h)\ge 0, we have Rrob(h)≥Rnat(h)\mathcal{R}_{\mathrm{rob}}(h)\ge\mathcal{R}_{\mathrm{nat}}(h): robust accuracy is upper-bounded by clean accuracy, and reducing the robust risk requires reducing the natural risk as well as the boundary risk. The clean degradation caused by replay increases Rnat\mathcal{R}_{\mathrm{nat}}, and prototype alignment addresses this term. Higher clean accuracy does not by itself reduce Rbdy\mathcal{R}_{\mathrm{bdy}} (Tsipras et al., 2019), which is why prototype alignment is paired with replay rather than replacing it. This trade-off, well documented for adversarial training (Zhang et al., 2019; Tsipras et al., 2019; Raghunathan et al., 2020), is visible in our ablation: adding the prototype-alignment component, which includes its centroid warm-up (C1 → C2 in Table R2), raises clean accuracy by 3.7 pp and lowers AA by 1.6 pp.

The AA cost appears at the training radius. Because robust accuracy approaches clean accuracy as ε decreases, the clean gain carries over to weaker attacks. We evaluate the models, all trained at ε = 8/255, at smaller radii:

\hline \text{Model} & \text{Clean} & \text{AA (8/255)} & \text{AA (4/255)} & \text{AA (2/255)} \\\\ \hline \text{C1 (replay)} & 80.58 & 43.83 & 62.79 & 71.40 \\\\ \text{C2 (+ warm-up / prototype alignment)} & 84.25 & 42.28 & 64.69 & 74.69 \\\\ \hline \text{FedR}^2\text{DG without prototype loss } (\lambda=0) & 82.53 & 45.75 & 64.23 & 73.61 \\\\ \text{FedR}^2\text{DG} & 83.97 & 45.59 & 64.89 & 75.09 \\\\ \hline \end{array}$$ The two pairs show the same pattern. Adding warm-up and prototype alignment together (C1 → C2) lowers AA by 1.6 pp at 8/255 but raises it by 1.9 and 3.3 pp at 4/255 and 2/255. Adding the prototype loss alone with warm-up fixed (FedR²DG without vs. with the prototype loss) changes AA by −0.2 pp at 8/255 and raises it at 4/255 and 2/255 in all three defended-client cases (by 0.7 and 1.5 pp on average). In both pairs, the AA gain at 2/255 is close to the clean gain (3.3 vs. 3.7 pp and 1.5 vs. 1.4 pp). *References.* H. Zhang, Y. Yu, J. Jiao, E. Xing, L. El Ghaoui, and M. Jordan. Theoretically principled trade-off between robustness and accuracy. ICML, 2019. D. Tsipras, S. Santurkar, L. Engstrom, A. Turner, and A. Madry. Robustness may be at odds with accuracy. ICLR, 2019. A. Raghunathan, S. M. Xie, F. Yang, J. Duchi, and P. Liang. Understanding and mitigating the tradeoff between robustness and accuracy. ICML, 2020. **Q4. FDG baselines** We add FedSR, FedDG-GA, and FedOMG under the same selective defense (Table R1). They improve clean accuracy but lower mean robust accuracy below SR-FedAvg (AA 10.4%, 20.4%, and 7.4%, respectively, vs. 27.2%), whereas FedR²DG reaches 45.6%. For FedSR, which has no adversarial training, we generate attacks and evaluate on the deterministic mean representation, so that training-time attacks target the evaluated model; training keeps its stochastic objective. FedDG-GA uses the official PACS cross-validation split for its generalization-gap estimate, following its implementation.
Reviewer_sJzcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates Selective Federated Robust Domain Generalization (SFRDG), aiming to reconcile adversarial robustness and domain generalization under heterogeneous client defense budgets. The authors first derive a held-out robust-risk bound that accounts for source-domain robust fitting, federated training drift, class-conditional source coverage, label-prior mismatch, and a conditional residual. Based on this analysis, they propose FedR²DG, which integrates budget-aware local adversarial training, EMA-based parameter tethering, global class-prototype alignment, and aggregation reweighting. Experiments on PACS, OfficeHome, and mini-DomainNet50 demonstrate improvements in adversarial robustness and the standard–robust accuracy trade-off across different perturbation budgets.

Critical strengths

1.The authors derive a held-out robust-risk bound and analyze several factors affecting robust generalization, including source-domain robust fitting, client drift, class-conditional distributions, and label-prior mismatch. This analysis provides theoretical motivation for the proposed framework. 2.The authors evaluate the proposed method on three benchmarks under different adversarial perturbation budgets using AutoAttack. The results demonstrate notable improvements in robust accuracy across most settings while largely preserving standard classification performance.

Critical weaknesses
  1. The proposed framework mainly integrates existing techniques, including adversarial training, EMA-based parameter tethering, prototype alignment, and aggregation strategies, with limited methodological innovation. 2.The fairness of the experimental comparisons requires further investigation. In the main experiments, the weakly defended clients of baselines such as SR-FedAvg and SR-FedProx primarily perform clean training, whereas those in FedR²DG additionally employ budgeted adversarial replay. Therefore, the performance improvements may partly result from involving more clients in adversarial training rather than from the other proposed mechanisms. 3.The current experiments mainly investigate the sensitivity to hyperparameters such as the prototype loss weight, temperature, and centroid update interval. However, the individual contributions of adversarial replay, EMA-based parameter tethering, prototype alignment, and aggregation reweighting are not systematically evaluated. More comprehensive component-wise ablation studies are needed to demonstrate the effectiveness and necessity of each design. 4.The evaluation under heterogeneous defense configurations is limited. The experiments primarily consider a fixed setting with three source clients, only one of which performs full adversarial training. However, different numbers of clients, defense ratios, and computational budget distributions may significantly affect performance. Additional experiments under more diverse federated configurations would help establish the generalizability of the proposed framework. 5.The motivation and distinct challenges of the proposed task are not sufficiently clarified. Although the paper introduces Selective Federated Robust Domain Generalization (SFRDG), it does not sufficiently explain the unique challenges of this setting compared with conventional federated domain generalization and federated adversarial learning, or why existing methods cannot be directly applied.
Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer sJzc

We thank the reviewer for the constructive comments.

W1. Integration of existing techniques

We agree that individual building blocks exist; our contribution lies in the SFRDG formulation and analysis and in adapting these blocks to selective defense (General Response, Positioning). Table R1 shows that existing FL, FL+AT, and FL+DG methods, each equipped with the same selective adversarial training, do not transfer robustness reliably to the unseen domain: every baseline falls to at most 18.6% AA when Sketch is the defended client. FedR²DG addresses this with proxy adversarial training on weakly defended clients at the per-round cost of clean training and recovers the resulting clean degradation through prototype alignment. Table R2 shows the contribution of each part.

W2. Fairness of the comparison

Replay is part of our method by design rather than a confound: it lets weakly defended clients contribute robustness at the same per-round cost as clean training (7.89 TMACs in both methods, Table 2).

Our design also differs from partial federated AT such as SFAT*, which keeps adversarially trained and standard-training clients as separate groups and aggregates them hierarchically. In FedR²DG, every client takes part in robust training within its own budget, so all clients are aggregated by a single weighted average (Eq. (12)) without being separated by defense type. Involving all clients in robust training through this unified framework is therefore the intended design and part of our contribution (General Response, Positioning (ii)).

To isolate the effect of replay, we evaluate SR-FedAvg with the identical replay schedule and all other components disabled (C1 in Table R2). Replay alone raises AA from 27.17% to 43.83%, and the remaining components add +3.4 pp clean accuracy, +1.8 pp AA, and +2.3 H on top of it (C1 → C4). We also give the baselines' weakly defended clients the same per-round budget for PGD-10 training:

\hline \text{Weakly defended client (per-round budget of 20 clean updates)} & \text{Clean} & \text{AA} & \text{H} \\\\ \hline \text{SR-FedAvg, 2 PGD-10 updates} & 79.92 & 38.06 & 49.44 \\\\ \text{SR-FedAvg, 10 clean updates + 1 PGD-10 update} & 82.97 & 42.37 & 55.79 \\\\ \text{SR-FedAvg + replay (C1)} & 80.58 & 43.83 & 56.60 \\\\ \text{FedR}^2\text{DG}\text{ (C4)} & 83.97 & 45.59 & 58.89 \\\\ \hline \end{array}$$ One PGD-10 update costs about 11 clean updates (Table 2: 86.79 vs. 7.89 TMACs per 20 updates), so both PGD allocations use slightly more compute (1.05–1.1×) than replay. With equal or less compute, replay alone (C1) already outperforms both PGD allocations, and FedR²DG is the strongest. **W3. Component-wise ablation** Please see Table R2 in the General Response. Replay provides most of the robustness; warm-up and prototype alignment recover clean accuracy while largely preserving robustness; √n weighting (over data-proportional weighting) and EMA tethering each improve the H-score. **W4. Diverse federated configurations** Please see Table R4 in the General Response, where we add corrupted copies of the PACS source domains as additional clients to obtain 6 or 9 source clients and defense ratios from 1/9 to 1/3. In every configuration, FedR²DG achieves the highest AA and H-score, exceeding the second-best method by 15.4–19.5 pp AA (17.1 pp on average). Its AA stays at 45.7–46.9% across all configurations, whereas SR-FedAvg drops from 27.6–28.1% at a defense ratio of 1/3 to 19.1% at 1/9. All weakly defended clients share the same per-round budget in these experiments; heterogeneous budgets among them are left for future work, which we will state in the limitations. We will report the per-case results (clean accuracy, AA, and H-score for each defended domain) for these five configurations on PACS-P in the appendix of the revision. **W5. Distinct challenges of SFRDG** SFRDG arises when institutions in a cross-silo federation differ in compute, so only some can afford full AT, and the model must serve a new institution that appears only at deployment (Section 1). SFRDG couples two difficulties that are not addressed together by FL+DG or FL+AT methods. First, robustness from a few defended clients transfers unevenly to the unseen domain: SR-FedAvg reaches 45.3% AA when Art is the defended client but 2.5% when Sketch is. Second, the mechanisms of each line do not resolve this: FL+DG methods improve clean accuracy but lower mean robust accuracy below SR-FedAvg, and SR-FRP, whose BN-based propagation cannot use target-domain statistics in this setting, falls to 6.6% AA (Table R1). FedR²DG addresses both: replay lets every weakly defended client fit robustness on its own domain, and prototype alignment preserves the clean class structure. It achieves the highest H-score in every defended-client case.
AuthorsOpenReview ↗

Follow-up for Reviewer sJzc

Dear Reviewer sJzc,

Thank you again for your review. We have uploaded a revised PDF (changes are marked in blue and summarized in General Response 2/2), which adds the component ablation (Section 5.3, Table 2), the equal-budget comparison (Appendix E.3), results with 6 or 9 clients and lower defense ratios (Appendix E.4), and heterogeneous budgets among weakly defended clients as future work (Section 6). Could you let us know whether any of your concerns remain unresolved? We would be glad to address them before the discussion period ends.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers (1/2)

We thank all reviewers for their careful reading. We summarize our positioning and the new experiments here; individual responses refer to these tables.

Positioning

To our knowledge, no prior method jointly addresses federated learning, unseen-domain generalization, and ℓ∞ adversarial robustness under heterogeneous defense budgets. FL+DG methods (FedSR, FedDG-GA, FedOMG) target clean generalization, and FL+AT methods (FRP, SFAT*) target the robustness of participating clients. Our contributions are (i) the SFRDG formulation and a held-out robust-risk bound that separates what a source-only federation can control from what it cannot; (ii) an adaptation of Free-AT (Shafahi et al., 2019) so that weakly defended clients perform proxy adversarial training at the per-round cost of clean training; and (iii) centroid-based prototype alignment that recovers the clean-accuracy degradation caused by this robust training. We do not claim that the bound is tight on the unseen domain.

Scope of new experiments

All new experiments use PACS with Photo held out, the three defended-client choices (Art, Cartoon, Sketch), ε=8/255, and a single seed, following the protocol of Figure 2, unless noted otherwise. Robust accuracy (AA) is measured with APGD-CE and APGD-T from the official AutoAttack implementation; the full configuration is given in our response to Reviewer djSo (W3/Q3). H is the harmonic mean of clean accuracy and AA computed for each defended-client case and then averaged, so it differs from the harmonic mean of the averaged Clean and AA columns.

Table R1. Comparison with FL, FL+AT, and FL+DG baselines (mean over three defended-client choices; last column = AA when Sketch is the defended client). All baselines use the same selective defense: the defended client runs PGD-10 AT, and weakly defended clients train with clean CE. FedDG-GA, SFAT*, and FedOMG are combined with FedAvg local training, following their official implementations.

\hline \text{Method} & \text{Type} & \text{Clean} & \text{AA} & \text{H} & \text{AA (def. S)} \\\\ \hline \text{Robust ERM (single client)} & \text{–} & 64.51 & 29.90 & 38.02 & 1.44 \\\\ \text{SR-FedAvg} & \text{FL} & 86.35 & 27.17 & 37.29 & 2.46 \\\\ \text{SR-FedProx} & \text{FL} & 86.59 & 27.42 & 37.66 & 2.81 \\\\ \text{SR-FRP} & \text{FL+AT} & 90.26 & 6.55 & 11.99 & 2.40 \\\\ \text{SR-SFAT*} & \text{FL+AT} & 87.84 & 28.64 & 38.76 & 2.87 \\\\ \text{SR-FedDG-GA} & \text{FL+DG} & 90.96 & 20.44 & 30.00 & 1.44 \\\\ \text{SR-FedOMG} & \text{FL+DG} & 87.54 & 7.35 & 13.28 & 1.86 \\\\ \text{SR-FedSR} & \text{FL+DG} & 91.16 & 10.44 & 18.17 & 18.56 \\\\ \mathbf{FedR}^{\mathbf{2}}\mathbf{DG} & & 83.97 & \mathbf{45.59} & \mathbf{58.89} & \mathbf{38.68} \\\\ \hline \end{array}$$ FedR²DG achieves the highest H-score in every defended-client case. As Table R1 shows, FL+DG methods improve clean accuracy but lower mean robust accuracy below SR-FedAvg, so their generalization mechanisms do not preserve the robustness contributed by the defended client. **Table R2. Cumulative component ablation** (mean over three defended-client choices; C2 adds warm-up together with prototype alignment because warm-up initializes the class centroids used by the prototype loss). $$\begin{array}{lcccc} \hline \text{Step} & \text{Added component} & \text{Clean} & \text{AA} & \text{H} \\\\ \hline \text{C0} & \text{SR-FedAvg} & 86.35 & 27.17 & 37.29 \\\\ \text{C1} & \text{+ replay (adapted Free-AT)} & 80.58 & 43.83 & 56.60 \\\\ \text{C2} & \text{+ warm-up / prototype alignment} & 84.25 & 42.28 & 56.11 \\\\ \text{C3} & \text{+ }\sqrt{n}\text{ weighting} & 85.31 & 43.97 & 57.78 \\\\ \text{C4} & \text{+ EMA tethering (= }\text{FedR}^2\text{DG}\text{)} & 83.97 & 45.59 & 58.89 \\\\ \hline \end{array}$$ Replay provides most of the robustness (AA +16.7 pp over SR-FedAvg) but lowers clean accuracy. Adding warm-up and prototype alignment then recovers clean accuracy (+3.7 pp) at a small robustness cost (AA −1.6 pp), consistent with its intended role. √n weighting (over data-proportional weighting) and EMA tethering each raise both AA and H (AA +1.7 and +1.6 pp; H +1.7 and +1.1). **Table R3. All-clients-AT upper bound.** $$\begin{array}{lcccc} \hline \text{Setting} & \text{Clean} & \text{AA} & \text{H} & \text{Weakly defended client cost per round (TMACs)} \\\\ \hline \text{SR-FedAvg} & 86.35 & 27.17 & 37.29 & 7.89 \\\\ \mathbf{FedR}^{\mathbf{2}}\mathbf{DG} & 83.97 & 45.59 & 58.89 & 7.89 \\\\ \text{FedAvg, PGD-10 AT on every client} & 80.96 & 49.88 & 61.73 & 86.79 \\\\ \hline \end{array}$$ FedR²DG recovers about 88% of the H-score gain from SR-FedAvg to this upper bound, while each weakly defended client spends about 1/11 of the per-round cost of PGD-10 AT.
AuthorsOpenReview ↗

General Response to All Reviewers (2/2)

Table R4. More source clients and lower defense ratios (mean AA on the clean held-out target over the three choices of defended domain; Robust ERM is trained only on the defended clients). Standard DG benchmarks contain only a few domains (four in PACS and OfficeHome), and federated DG typically assigns one source domain to each client, so leave-one-domain-out evaluation leaves only three source clients. To vary the number of clients and the defense ratio while keeping the held-out target unseen, we add corrupted copies of each source domain as additional clients (Gaussian noise and defocus blur, severity 1; Hendrycks & Dietterich, 2019). With six clients, each source domain contributes its clean and Gaussian-noise copies; with nine, also its defocus-blur copy. The defended clients are the clean copy of one defended domain, plus its Gaussian-noise copy when two clients are defended and also its defocus-blur copy when three are.

\hline \text{Sources / defended (ratio)} & \text{Robust ERM} & \text{SR-FedAvg} & \text{SR-FedProx} & \text{SR-FRP} & \text{SR-SFAT*} & \text{FedR}^2\text{DG} \\\\ \hline \text{6 / 2 (1/3)} & 29.90 & 27.60 & 27.56 & 7.64 & 30.10 & \mathbf{46.31} \\\\ \text{9 / 3 (1/3)} & 29.30 & 28.06 & 28.50 & 4.59 & 30.58 & \mathbf{45.95} \\\\ \text{6 / 1 (1/6)} & 27.33 & 25.13 & 24.97 & 3.47 & 27.10 & \mathbf{46.87} \\\\ \text{9 / 2 (2/9)} & 30.06 & 24.93 & 24.89 & 3.10 & 27.01 & \mathbf{46.59} \\\\ \text{9 / 1 (1/9)} & 28.02 & 19.08 & 18.84 & 1.60 & 21.54 & \mathbf{45.69} \\\\ \hline \end{array}$$ In every configuration, FedR²DG achieves the highest AA and H-score, exceeding the second-best method by 15.4–19.5 pp AA (17.1 pp on average; H 58.2–60.0 vs. at most 41.5 for any baseline). **Changes in the revision** (marked in blue in the revised PDF) - **Figure 1:** Updated to show the components of FedR²DG together with the problem setting. (djSo W6) - **Section 2 (Related Work):** Added SFAT, SFAT\*, and FAML to federated adversarial training, and FedSR and FedOMG to federated DG. (djSo W1/Q1 and W4/Q4, Bvvy Q4) - **Section 5.2 and Table 1:** Stated the threat model (white-box ℓ∞ evasion attacks on the held-out domain at the training radius) and clarified in the text and the Table 1 caption that AA denotes APGD-CE + APGD-T. (djSo W3/Q3) - **Section 5.3 and Table 2:** Added the cumulative component ablation (Table R2) and the upper bound with PGD-10 AT on every client (Table R3). (sJzc W3, Bvvy W2, g5rc W1/W2, djSo W4/Q4) - **Section 5.3:** Added a paragraph on more source clients and lower defense ratios (Table R4, details in Appendix E.4) and a paragraph on EMA tethering and client drift (details in Appendix E.1). (sJzc W4, g5rc W1/W2, Bvvy Q1) - **Section 5.4:** Clarified that Table 3 counts local computation only and quantified the extra communication of centroid refreshes. (g5rc Q3) - **Section 6 (Limitations):** Noted that heterogeneous budgets among weakly defended clients remain future work. (sJzc W4) - **Appendix E (new, Additional Analyses):** - E.1, Table 24 and Figure 3: Sensitivity to γ_ema and parameter drift over training rounds. (Bvvy Q1) - E.2, Table 25: Aggregation weighting and the label-prior radius. (Bvvy Q2) - E.3, Table 26: Equal-budget comparison for weakly defended clients. (sJzc W2, g5rc W3/Q1, djSo W5/Q5) - E.4, Tables 27–28: Mean and per-case results with 6 or 9 source clients and defense ratios from 1/9 to 1/3, including SR-SFAT\*. (sJzc W4, djSo W4/Q4) - **Appendix F.3.1 and Table 29:** Exact AutoAttack configuration. (djSo W3/Q3) Table numbers in our individual responses refer to the original submission. In the revision, Table 2 is now Table 3, and the appendix tables are shifted by one (e.g., Table 13 is now Table 14).
AuthorsOpenReview ↗

Summary of the Discussion and Revision

We thank the reviewers and the Area Chair. Tables R1–R4 are in the General Response, and revisions are in blue.

1. Contributions acknowledged by the reviewers

  • Practical problem and motivation. The reviewers found the setting important and practically relevant (Bvvy, g5rc), with a sound motivation (djSo).
  • Theoretical analysis. The held-out robust-risk bound clarifies why source-side robustness alone is insufficient and motivates the method (sJzc, Bvvy), with a clear formulation (g5rc) and a rigorous, traceable derivation (djSo).
  • Experiments and presentation. The experiments are comprehensive (g5rc) and show notable improvements in robust accuracy across most settings while largely preserving standard accuracy (sJzc). The paper is well organized and reproducible (djSo).

2. Concerns and how we addressed them

  • Novelty and distinct challenges (sJzc W1/W5, Bvvy W1, djSo W1/W2). To our knowledge, no prior method jointly addresses FL, unseen-domain generalization, and adversarial robustness under heterogeneous defense budgets (General Response, Positioning). The central difficulty of this setting is that robustness from a few defended clients transfers unevenly to the unseen domain (SR-FedAvg: 45.3% AA with Art defended vs. 2.5% with Sketch), and existing FL+AT and FL+DG methods do not resolve it (Table R1). Section 2 now covers SFAT, SFAT*, FAML, FedSR, and FedOMG.
  • Additional baselines (Bvvy Q4, djSo W4). Table R1 adds FL+AT (SFAT*) and FL+DG (FedSR, FedDG-GA, FedOMG) baselines under the same selective defense. With Sketch defended, every baseline falls to at most 18.6% AA, whereas FedR²DG reaches 38.7%.
  • Component ablation and upper bound (sJzc W3, Bvvy W2, g5rc W1/W2, djSo W4). The cumulative ablation and the upper bound with PGD-10 AT on every client (Tables R2–R3) are added to Section 5.3. FedR²DG recovers about 88% of the H-score gain toward this bound at about 1/11 of the weakly defended clients' per-round cost.
  • Fairness of the comparison (sJzc W2, g5rc W3/Q1/Q2, djSo W5). Under an equal budget, replay outperforms PGD-10 training for weakly defended clients (C1, Appendix E.3), and warm-up mainly raises clean accuracy, not robustness.
  • More clients and defense ratios (sJzc W4, g5rc W1). With 6 or 9 source clients and defense ratios of 1/9–1/3, FedR²DG achieves the highest AA and H-score in every configuration (Section 5.3, Appendix E.4).
  • Threat model and AutoAttack configuration (djSo W3). Stated in Section 5.2 and Appendix F.3.1 (Table 29).
  • Drift, weighting, and prototype alignment (Bvvy Q1–Q3). Appendix E.1–E.2 add γ_ema sensitivity with parameter drift and weighting schemes, and Table R2 shows that prototype alignment recovers clean accuracy.
  • Cost scope and figures (g5rc Q3, djSo W6). Section 5.4 clarifies the cost scope and communication overhead. Figure 1 now shows the framework, and Figure 3 shows parameter drift.

3. Remaining points

Reviewer Bvvy asked how much sensitivity to the defended-client choice remains with replay. Across all 12 held-out targets, the AA spread is smaller for FedR²DG in every target (4.2 pp vs. 13.7 pp for SR-FedAvg on average).

Reviewer djSo noted that the additions improved the paper and raised three remaining points, which concern the completeness of the evaluation rather than the main claims.

(1) For FAB-T and Square, the same protocol is applied to every method, and prior work, including the original AutoAttack paper (Croce & Hein, 2020) and Debenedetti et al. (2023), shows that these attacks rarely change the result after APGD-CE and APGD-T. Other studies also use reduced attack sets for efficiency (Singh et al., 2023, Fort & Lakshminarayanan, 2024).

(2) For federated Free-AT, replay indeed provides most of the robustness gain, but the remaining components add 3.4 pp clean accuracy, 1.8 pp AA, and 2.3 H on top of it (C1 → C4), so our contribution does not rest on the Free-AT adaptation alone. Unlike SFAT*, which aggregates adversarially and standard-trained clients hierarchically as separate groups, FedR²DG intentionally aggregates all clients in one weighted average. Our illustrative example shows that carrying Free-AT's perturbations across aggregation can weaken the next global model's robustness, and we will add a direct comparison with an unmodified federated Free-AT to the final version.

(3) For the equal-budget comparison with one seed and domain, our main results cover three seeds and 108 combinations over 12 held-out domains, where FedR²DG ranks first by robust accuracy in 105.

Therefore, we believe these are minor points that do not affect the main contributions. djSo's novelty concern is addressed in the first item of Part 2.

4. Reviewer without a response

Reviewer sJzc has not yet responded to our rebuttal. The concerns marked with sJzc in Part 2 are addressed in our responses and reflected in the revised version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.