acceptodds
Withdrawn from ICLR 2027

FUSE: Fairness-aware Update Screening against Byzantine Exploitation in -Fair Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Client-level fair federated learning, in the form of -fair federated learning (-FFL), agnostic federated learning, or tilted empirical risk minimization, has emerged as an attractive solution to equalize utility across heterogeneous clients. This paradigm grants high-loss clients greater relative influence, while a central server aggregates their loss-weighted updates into a single shared model. Yet, the Byzantine robustness of -fair federated learning has been largely neglected, leaving the server unable to tell legitimate fairness-induced amplification from malicious update scaling. To address this critical gap, we formalize the problem of Byzantine-robust -fair federated learning under an update-only interface, in which the server observes nothing but the fairness-scaled client updates. Our analysis under shared class-conditional label skew shows that fairness scaling preserves the nonnegative class-mixture structure of honest updates, while replacing their unit-sum constraint by a client-dependent radial scale. Moving beyond magnitude-based screening, we show that this warped geometry supplies two complementary tests, one radial and one structural, that a single norm cannot provide. We then introduce FUSE, a structure-aware robust aggregator that uses a small trusted labeled dataset for coefficient recovery, scale and cone-consistency screening, and post-screening step-size control. Across four benchmarks and seven Byzantine attacks, FUSE improves seed-wise worst-attack accuracy over the strongest baseline on each benchmark by 3.9–30.1 percentage points, while largely preserving the influence of the high-loss honest clients that fairness is meant to protect.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fair Federated Learning, Byzantine Robustness, Client-Level Fairness, Robust Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CaYnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The present paper explores Byzantine robustness in q-fair federated learning, whereby each client contributes its update scaled by its loss. Since both of these are combined, the server is unable to distinguish honest clients that have high losses from adversaries scaling their updates. The authors demonstrate that under the assumption of label skew with shared class-conditionals, the honest clients' updates continue being nonnegative combinations of class directions, but this time with weights summing up to the fairness parameter, such that a cone is formed rather than a simplex. The FUSE aggregator makes use of a small labeled dataset to construct class reference directions, recover the coefficients of the updates via nonnegative least squares, filter based on the mass of the coefficients and the cone residual, and normalize the step sizes using estimated losses. Both recovery and exact-screening guarantees are established. On four benchmark datasets and seven attacks, FUSE achieves an increase in accuracy on worst-attack performance of 3.9-30.1 points while retaining most high-loss honest clients.

Critical strengths
  • The main point of the author's work is that when you scale for fairness, you are able to keep honest updates as non-negative class mixtures. However, you change the coefficient sum of those honest updates. When you do this, you turn what would have been a simplex into a cone. This is not simply stated by the authors. The authors are able to show that when using the checkpoint, client, and trusted basis as fixed components, the cone residual remains at 0.178 on MNIST while the simplex residual increases from 0.185 to 1.077. Additionally, when the authors swapped out only the cone score with a simplex score within the same pipeline, the Byzantine recall for Min-Max and Min-Sum fell from 1.00 down to approximately 0.6. This demonstrates that the improvements seen were due to the proposed idea itself, and is why I feel that the author's contribution is valid.

  • Robustness without compromising on fairness, backed up with thorough evaluation. On each of the four datasets, FUSE has the highest worst attack accuracy (3.9 to 30.1 percent higher than the most robust benchmark) when evaluated with a conservative worst attack metric based upon seed-wise worst attacks across five different seeds and seven different types of attacks. In addition to being the most robust approach to Byzantine attacks, the authors also show which clients were rejected by the system. While FUSE rejected only 0-9.7 percent of those clients that had the largest loss during training, and were not malicious, the norm-based filtering approaches rejected 22-36% of those same clients. However, FUSE was able to reject 85.7-99.8 percent of all submissions from malicious clients. These findings provide strong evidence in support of the author's primary claim, and the transparency of the authors' reporting (including showing failed runs, etc.) helps build confidence in the authors' results.

Critical weaknesses
  • The practical significance of the fairness being preserved is not argued for, and the comparison could be in favor of FUSE. Fairness is evaluated by rejections rather than by improvement of q-fair training on the weak clients. The authors' own figures cast doubt on that. For MNIST, accuracy for the worst-10% group decreases as q increases (86.21 when q = 0.5 and 84.85 when q = 2). Also, the q=0 simplex screening gives 96.55/92.13 clean and 93.95/89.26 when q = 1 and robustness against tested attacks. There is no baseline that removes fairness scaling and applies an existing robust defense, so it remains unclear whether the defense of q-FFL is worth it.

  • Second, all approaches seem to employ a single learning rate for each dataset when applying loss-scaling updates. Some baselines are unstable even without attack (q-FFL on EMNIST: 63.79 ± 34.5), and FUSE comes with step-size control out-of-the-box. It is significant since it concerns the motivation and the main comparison.

  • The evidence backing the central arguments is less convincing than claimed. Exact recovery is only guaranteed by the theory when the number of Byzantines is known, and each attack is far apart, using metrics on the server side that are not available. No experiment has ever verified these preconditions on their benchmark data sets. Moreover, there is no proof of convergence and no bound when the screening step fails, which is where the experiments take place.

  • The analysis of the mechanism, aimed at verifying the cone geometry benefits, also contains incongruities between the text and the tables. In particular, the AGNews ablation study mentions 53.34/38.33 for the full FUSE mechanism and characterizes the loss normalization as a trade-off, but Table 16 gives 79.45/62.86 for the normalized version and 35.92/25.22 for the non-normalized one. According to Table 16, the version without scale screening outperforms full FUSE despite the claim that this stage is never harmful.

  • IPM recall of the cone score 0.69 in Section 6.2, 0.62 in Table 20 and 0.59 in Table 17. Given that this analysis involves just one seed, the mentioned inaccuracies undermine my confidence in the conclusions about the components' performance.

Questions for authors
  • As shown in Table 19, for MNIST, the q = 0 run gets 96.55/92.13 compared to 93.95/89.26 for q = 1, while Table 23 shows that worst-10% accuracy decreases with q. Is it the difference in the step size because the losses are less than one, or do q-fair optimization methods not favor the least well-performing clients here? Do you have any prior results where q > 0 helps improve worst-10% accuracy or reduces client accuracies' variance compared to q = 0?

  • Have you used different learning rates for each method, or did you set one in Table 4 for everyone? Some baselines appear quite unstable without an attack (for instance, q-FFL on EMNIST with 63.79 ± 34.5). Did they get some step size control like Lipschitz normalization in q-FedAvg?

  • The numbers mentioned in E.1 are 53.34/38.33 for full FUSE and 55.45/40.77 when not using loss normalization, whereas Table 16 gives 79.45/62.86 and 35.92/25.22. Also, the text mentions 1000 rounds, whereas the caption states 2000. Which are the right numbers? If Table 16 is correct, how to explain the fact that the variant without the scale screen (80.44/66.99) performs better than the full FUSE given the claim that this phase never decreases the worst-attack accuracy?

  • The recall of the cone-score on the MNIST dataset under IPM in Section 6.2 is 0.69, 0.62 in Table 20, and 59.4% in Table 17. Are these separate runs or different pooling windows? Please specify which run demonstrates the 2.44 gain.

  • From your logs, give the measured honest radial spread r_H and cone variation, as well as the frequency at which log κ > 2r_H and the other separation conditions hold for each of the attacks?

1 comment

Thanks for the careful review. We answer the five questions below.

1. Does q>0q>0 actually improve fairness here?

The original q-FFL paper does show that positive qq can improve low-performing clients, but qq is tuned per dataset rather than fixed to 1. Li et al. (2020, Table 2/App. F.1) report:

Datasetq=0q=0 W10 / Var.tuned q>0q>0 W10 / Var.
Synthetic18.8 / 72431.1 / 472 (q=1q=1)
Vehicle43.0 / 29169.9 / 48 (q=5q=5)
Sent14015.9 / 69723.0 / 509 (q=1q=1)
Shakespeare39.7 / 8242.1 / 54 (q=.001q=.001)

Thus q-FFL motivates giving high-loss clients more relative influence, but it does not imply that q=1q=1 must improve W10 on every dataset.

We also reran q=0/1/2q=0/1/2 with the same tuning budget for each qq: 3 LRs x 2 clean tuning seeds, then 3 fresh confirmation seeds.

DataqqW10Client variance
MNIST0 / 1 / 294.63 / 91.52 / 88.912.25 / 4.79 / 6.75
CIFAR0 / 1 / 278.82 / 78.26 / 76.5416.51 / 15.82 / 15.22

So in our setting, q=1q=1 does not give a clear W10 gain. On CIFAR it slightly reduces variance, while on MNIST it does not. We will therefore narrow the claim: FUSE preserves the intended high-loss weighting when q>0q>0 is used; we do not claim that q=1q=1 itself always improves fairness.

The reviewer is also right that step size changes with qq. Our q-FedAvg diagnostic gives effective α=0.090/0.00524/0.00294\alpha=0.090/0.00524/0.00294 for CIFAR q=0/1/2q=0/1/2 and 0.300/0.140/0.0830.300/0.140/0.083 on MNIST. This suggests under-optimization may contribute, but it is not a causal proof.

2. Were learning rates and step-size control compared fairly?

The original main table used one LR per dataset, so we added a clean-tuned comparison. Every method/qq gets 3 LR candidates x 2 validation seeds; the selected setting is frozen before 3 fresh test seeds. BOBA/FLTrust use native q=0q=0 and H-nobs uses q=1q=1.

We also separated screening from normalization with a fixed-LR 2×22\times2 test:

DataSelectorMean scalar W10Normalized W10
CIFARnone0.00.0
CIFARFUSE47.670.0
EMNISTnone0.00.0
EMNISTFUSE42.870.8

Thus normalization helps once bad updates are screened, but normalization alone does not provide robustness. We will add the independently tuned native-baseline results and make the LR protocol explicit.

3. Why do the AGNews numbers differ?

The manuscript mixes two diagnostic protocols and did not label them clearly enough.

SourceProtocolFull FUSEno loss norm.
App. E.1 textsingle-seed, 1000-round factorial diagnostic53.34/38.3355.45/40.77
Table 16single-seed, 2000-round AGNews endpoint79.45/62.8635.92/25.22

The shorter 1000-round factorial was used to keep the full component grid affordable; AGNews's main horizon is 2000 rounds. The sentence in E.1 saying Tables 13–16 all use 1000 rounds is therefore misleading and will be corrected. These numbers should not have been compared as if they came from one protocol.

This also fixes the wording about the scale stage: we will not say it “never hurts.” Its effect is attack- and horizon-dependent; the ablation only supports its role against radial attacks.

4. Why are the IPM recalls 0.69, 0.62, and 59.4%?

These numbers come from different analyses, and one summary number was not synchronized:

LocationExperimentIPM recall
Table 17production FUSE, seed 42, pooled over 1000 rounds59.4%
Table 20matched simplex-vs-cone control, seed 42, pooled over 1000 rounds62%
Sec. 6.2 textsummary of the matched control69%

The 2.44-point gain is from the matched q=1q=1 training control in Tables 19–20: global accuracy is 89.91 with simplex screening and 92.35 with scale+cone, i.e. +2.44 points. Table 20's 62% recall is the corresponding screening statistic. Therefore the 0.69 in Sec. 6.2 is a stale summary value and will be corrected to 0.62. Table 17 remains separate because it reports the production FUSE run, not the matched simplex-control run.

5. Do the theorem conditions hold in training?

We agree this should be shown directly. We audited the round-level separation conditions over 3 seeds:

AttackCIFAR separationEMNIST separation
SignFlip59.0%84.5%
IPM71.9%86.6%
Gaussian100.0%100.0%
ALIE66.9%77.4%
Magnitude99.8%100.0%
Min-Max96.7%100.0%
Min-Sum99.2%100.0%

This also explains where the theorem is strong and where it is not: Magnitude/Min-Max/Min-Sum satisfy the separation conditions almost always, while SignFlip and IPM overlap more with honest updates. The theorem is therefore a conditional certificate, not a claim that its sufficient conditions hold every round. We will add this audit and report the corresponding honest radial spread and cone variation with it.

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

This paper focuses on the problem of Byzantine attacks in fair federated learning. In q-FFL, updates from clients with high loss are amplified to improve their training performance. However, this creates a vulnerability which malicious clients can also amplify their updates to attack the global model. This vulnerability leads difficult for the server to distinguish between a legitimate, fairness-weighted update and a malicious attack. Existing robust aggregation methods may inadvertently filter out valid updates from clients experiencing high loss. To tackle this problem, this paper proposes FUSE. The core idea is to leverage the geometric structure of client updates, rather than relying solely on their magnitude to identify malicious contributions. Specifically, this paper proposes that fairness weighting transforms the original simplex structure into a nonnegative cone. Therefore, they use a small amount of trusted data and a non-negative least squares approach to estimate the combination coefficients of the client updates. They then employ a two-stage process to check for anomalies in both the magnitude and the geometric structure of these updates. Finally, they utilize the estimated client losses to further refine the aggregation process.

Critical strengths
  1. The research problem and the motivation of this work is good. This paper focused on the gap of traditional methods often misidentify these legitimate updates as malicious attacks and propose the geometric structure of the updates to distinguish between benign clients and attackers which is kind novel.
  2. The paper did a lots of experiment which is relatively comprehansive which prove the idea claimed by this paper.
Critical weaknesses
  1. One of my major concern is the novelty, BOBA already apply updated geometric structures on the client side to detect Byzantine attacks, while FUSE primarily extends the original Simplex into a nonnegative cone, incorporating NNLS alongside a two-stage anomaly detection process. Although this modification is well-founded and yields strong experimental results, the overall approach is, from a technical standpoint is quite straightforward.
  2. In addtion, I have some concerns about FUSE examines only the classifier parameters across several vision datasets, the final aggregated result incorporates updates from the entire model. This allows an attacker to maintain normal behavior for the parameters being inspected while simultaneously launching attacks via other parameters. the primary experiments assume prior knowledge of the number of Byzantine clients setting f=4f=4 when attacks are present and f=0f=0 when they are absent. In real-world systems, however, the exact number of attackers is typically unknown.
Questions for authors
  1. This paper uses estimated client losses to adjust the global aggregation step size, yet they do not fully explain how this adjustment affects the original q-FFL optimization objective. While global normalization does not directly alter the relative weights among the retained clients, the filtering of clients which combined with variations in the step size at each round could still influence the final fairness and the convergence process. Could the authors provide further clarification on the conditions under which FUSE remains capable of optimizing the original q-FFL objective?
  2. FUSE constructs class-specific reference gradients using a small amount of trusted, labeled data on the server, while theoretically assuming that all clients share the same class-conditional distribution. In practical non-IID federated learning scenarios, however, one may encounter label skew, feature skew, and even concept shift. Although The paper certain instances of covariate shift, are these experiments sufficient to demonstrate that the reference gradient remains an accurate representation of updates from benign clients? In particular, when the server’s trusted data distribution differs from that of the clients, how do the authors distinguish whether the fitting error in NNLS stems from benign data heterogeneity or from a malicious attack? Furthermore, to what extent do the current theoretical guarantees cover such scenarios?
  3. In a pratical federated learning settings, how does the server obtain this information? Furthermore, if only an upper bound on the number of Byzantine clients is known rather than the exact count to what extent can FUSE still maintain fairness? Because, FUSE requires a pre-declared rejection budget (f). In the main experiments, a value of f=0f=0 is used when no attacks are present, while f=4f=4 is used when attacks occur effectively assuming prior knowledge of whether an attack is taking place and the number of attackers involved.
  4. I understand that the transition from a simplex to a nonnegative cone is the core innovation of the paper, but I’m not entirely certain how much of the final performance gain actually stems from this geometric shift.
  5. I would like to know why the authors believe that the results from the seven standard attacks are sufficient to support a stronger claim of Byzantine robustness. Given that an attacker might understand FUSE’s filtering mechanism and be able to design updates specifically to target it, what level of robustness guarantee does FUSE actually provide in such a scenario? Because the baselines are relatively old, only one from 2025.
  6. The table in appendix are hard to read and one table looks beyond the content margin. I think related work should be include in the main content instead of put all of the section in the appendix.
1 comment

Thanks for the review. We answer the six questions below.

1. Effect on the q-FFL objective.
Eq. 5 applies one positive scalar to the retained mean, so it changes step size, not direction. App. F.4, Prop. 1 already gives the descent condition: if JqJ_q is smooth, the retained full-model update aligns with −∇Jq-\nabla J_q, and the step is below the stated bound, then Jq(wt+1)<Jq(wt)J_q(w_{t+1})<J_q(w_t). Thus we claim conditional descent on the original q-FFL objective, not unconditional convergence.

2. Trusted-data mismatch.
Assumption 2 already states that the theory does not cover arbitrary changes in class-conditionals. App. E.13 tests covariate shift using an unshifted trusted set: honest FPR stays at 1.4–2.0%, retained fair mass at 95–96%, and worst-attack accuracy stays within about 1 point of shifted-clean. Changing trusted data from 5 to 50 samples/class changes worst-attack accuracy by at most 1.5 points. This supports the tested shifts, not arbitrary concept shift; if an honest client leaves the trusted cone, NNLS error alone cannot tell benign shift from attack.

3. Unknown attacker count ff.
As in BOBA, ff is a declared Byzantine upper bound, not known attacker identities or the exact count (Bao et al., 2024; Assumption 1). App. E.12 already tests f=4f=4 with only 0/1/2/3 actual attackers. With 1–3 attackers, accuracy stays within 0.7 points of the matched case. With 0 attackers, accuracy changes little, but retained fair mass falls to 59–62%. Thus a loose upper bound mainly costs fairness through extra rejection.

4. Novelty and the gain from cone geometry.
Our main novelty is not simply “simplex to cone.” The key finding is that q-fair weighting breaks BOBA's absolute Euclidean-distance criterion. BOBA ranks updates by distance to an affine subspace, then checks simplex coefficients. Under q-FFL, an honest update is scaled by FiqF_i^q, so a high-loss honest client can be pushed far from that affine simplex only because fairness gives it more weight. Thus, large absolute distance no longer means maliciousness (Sec. 3).

FUSE instead uses coefficient mass relative to honest scale for radial anomalies and a scale-normalized cone residual for structural anomalies (Sec. 4).

New rebuttal experiment: only the Stage-I statistic is changed:

DataStat.High-loss FPRFair massRecall
CIFARmass2.3899.1098.21
CIFARnorm10.7196.5393.45
EMNISTmass/norm0.0099.6499.01

Sec. 6.2 also isolates simplex vs. cone scoring. On IPM, recall rises 45%→69% while honest FPR drops 21.9%→12.4%. On Min-Max/Min-Sum, recall rises 56%/62%→100%/100%, while honest FPR drops 17.5%/15.3%→0.1%/0%. SignFlip is the counterexample where simplex is better. Table 3 shows the same end-to-end effect: removing cone screening drops Min-Max/Min-Sum from 79.28/78.71 to 36.71/38.94.

5. Robustness scope, classifier-only screening, and baselines.
We do not claim that seven standard attacks prove arbitrary robustness. App. E.14 already contains a filter-aware attack that stays inside the cone with plausible mass; CIFAR terminal accuracy drops from 82.4% clean to 65.6%. This matches Theorem 2: without radial or structural separation, FUSE has no certificate.

The classifier-only setting is a computational choice, not a requirement of FUSE; AGNews already screens all parameters. Full-space structure screening is expensive on ResNet-18, e.g., our full-space BOBA reproduction takes 30–47 GPU-hours/run (App. B.5). App. E.15 also directly tests the reviewer's attack: keep classifier coordinates benign and attack only uninspected coordinates. It causes collapse on CIFAR, MNIST, and EMNIST. Thus screening protects only observed coordinates, and we will state this boundary more clearly.

We agree most baselines are not recent. They were chosen by relevance: H-nobs is fairness-aware, BOBA geometry-based, FLTrust uses trusted server data, and FedAA (2025) is included. We did not add another 2025/2026 method during rebuttal.

We did add a native-setting comparison. BOBA/FLTrust use q=0q=0, each gets the same clean tuning budget (3 LRs x 2 tuning seeds), followed by 3 fresh seeds:

DataMethodW10 (Sign / IPM / Mag.)
MNISTFUSE93.65 / 92.47 / 91.73
FLTrust92.87 / 92.71 / 94.53
BOBA92.85 / 30.90 / 93.20
CIFARFUSE76.84 / 75.66 / 78.24
FLTrust75.99 / 75.99 / 72.33
BOBA56.18 / 55.70 / 59.02

The ranking is mixed, but BOBA still drops badly on IPM, so its earlier gap is not explained only by using q=1q=1.

6. Presentation.
We agree. We will fix the oversized appendix tables and move a short Related Work discussion into the main text.

Reviewer_rzL4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper considers the problem of Byzantine-robustness in q-fair FL. To solve this problem, the paper proposes FUSE, a structure-aware robust aggregation scheme that uses a trusted labeled dataset. The paper provides an empirical study to demonstrate the advantages of the proposed scheme.

Critical strengths
  1. The problem is important. Both fairness and robustness are important issues in FL, and their synthesis has received little attention.
  2. The empirical results demonstrate some advantage of the proposed scheme.
Critical weaknesses
  1. Lack of motivation. It is not clearly argued why existing robust aggregation schemes cannot be used to aggregate the uitu^t_i’s. In fact, it is not clear how the problem differs from standard Byzantine-robustness with the loss functions FiF_i simply substituted by FiqF_i^q.
  2. Unsupported claims and sparse technical details. Several claims are made in passing without concrete arguments or reasoning. For example: (i) the authors claim that magnitude-based schemes will be ineffective, but do not define what they mean by magnitude-based schemes, leaving the claim ambiguous; (ii) the paragraph on structure-aware robust aggregation in Line 67 starts with “Building on this geometry ...” without it being clear what “this geometry” refers to; (iii) it is not explained why it matters that standard aggregation breaks the unit-sum constraint for honest updates, nor how the proposed scheme satisfies this constraint. There are several such issues, and this list is not exhaustive.
  3. Reliance on a trusted dataset. The scheme assumes the server has access to a small validated labeled dataset, and the paper does not argue why the server does not simply optimize over this dataset, i.e., what is the benefit of federation in this setting in the first place? Moreover, the paper does not discuss how sensitive the algorithm is to the size of this dataset.
Questions for authors
  1. How does the problem differ from standard Byzantine-robustness applied to the losses FiqF_i^q, and why do existing robust aggregators fail on the uitu^t_i’s?
  2. What precisely is meant by “magnitude-based schemes,” and on what basis are they claimed to be ineffective here?
  3. If the server already has access to a small validated labeled dataset, what is the benefit of federation?
  4. Why does it matter that standard aggregation breaks the unit-sum constraint for honest updates, and how is this constraint satisfied by the proposed scheme?
  5. In Assumption 2, why is ℓc\ell_c bounded away from 00?
  6. How are the parameters κ\kappa and ϵ\epsilon chosen?

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

Related papers

Discussion (0)

Sign in to comment.