acceptodds
Under review as a conference paper at ICLR 2027

CARAT: Class-Aware Robust Aggregation for Model-Poisoning-Resilient Federated Learning

Abstract

Federated learning is vulnerable to model-poisoning attacks because the server must aggregate client updates without inspecting client data. Existing robust aggregation rules often rely on update geometry, coordinate-wise robust statistics, or a single trusted reference direction, but these signals can confuse malicious behavior with benign client heterogeneity under non-i.i.d. data. We propose CARAT, a class-aware robust aggregation rule that uses a small labeled server reference set to sample hidden, class-balanced tasks and score each candidate update by its task-level loss reduction. CARAT combines these hidden-task certificates with robust clipping, common-radius evaluation, coordinate-rank anomaly penalties, and capped-simplex weight optimization to favor updates that remain useful across semantic slices while limiting structurally extreme updates. On CIFAR-100/ResNet18 with malicious clients, two non-i.i.d. regimes, and four untargeted model-poisoning attacks, CARAT achieves the best average rank among the reported public baselines and improves final accuracy by – percentage points over the strongest non-CARAT baseline in each attack/heterogeneity setting. Within this scoped evaluation, the results support task-conditioned validation as a practical alternative to purely geometric or single-anchor trust signals when a modest labeled server reference set is available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated learning, model-poisoning attacks, robust aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bi44OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a method to tackle model-poisoning in federated learning, which is CARAT. CARAT uses a validation set on the server side to construct a class-aware set of representations and use them to construct certificates for later aggregation. The empirical results are extensive and support the claims.

Critical strengths
  • Good visualization
  • Extensive empirical results
Critical weaknesses
  • In Section 4.2, what is the rationale of clipping the update of each client. If the goal is to avoid dominant scale, simple standard normalization would be sufficient. The author should provide ablation studies to verify this design choice.
  • The design of certificate (Eq. 8) is based on the hypothesis that malicious updates might degrade performances of only a subset of tasks. Is there any evidence to show this hypothesis is correct empirically?
  • The design choices are very unclear to me. It is hard to understand why the author designed the method with complicated components without verification, e.g., coordinate-rank anomaly score or bounded weight learning. More experiments to verify them are necessary.
Questions for authors

See Weaknesses

2 comments

Rationale for clipping instead of normalization (Sec. 4.2). The two operations serve different purposes, and CARAT uses both.

  • Scoring. For the certificate, every update is rescaled to a common radius (Sec. 4.2), so its score depends on direction only (the directional-invariance result in Appendix A). This is the normalization the reviewer suggests.
  • Aggregation. Normalizing every update to the same norm would give a near-zero (uninformative) update the same influence as a full benign update, and would remove benign magnitude information under non-i.i.d. data.
  • What clipping does. MAD clipping only bounds the upper tail: it removes the advantage of inflated updates and leaves typical benign updates untouched. Together with the weight cap, it gives the aggregate norm bound in Appendix A.
  • Ablation (Table 4). Removing clipping lowers accuracy (46.26 vs 46.81 under ALIE, 46.79 vs 47.59 under FangAttack).

Evidence for the "subset of tasks" hypothesis behind Eq. 8. Section 4.2 now presents this as a design motivation. The softmin is a conservative aggregate: it focuses the weight optimization on the worst-served hidden task rather than the average, so harm on any single task lowers the objective. Empirically, the full rule removes aggressive attackers:

  • FangAttack updates obtain (near-)zero weight in late training at 20–40% malicious.
  • ALIE at 40% obtains (near-)zero weight in early training, while undefended averaging collapses.

Verification of the rank penalty and the bounded weights.

  • Rank penalty. In the per-round logs, FangAttack attackers receive rank penalties 2.8–17 robust (MAD) units higher than benign clients at 30–40% malicious. Removing the penalty lowers accuracy in Table 4 (46.36 vs 46.81, 46.96 vs 47.59).
  • Bounded weights. The cap 1/((1−ρ̂)n) keeps any client, including a mis-scored one, from dominating the aggregate (Appendix A). We now report the resulting malicious mass q_M (see the response to jAzV).

Every component improves on its removal in Table 4. The differences are modest at 20% malicious because in that regime even undefended averaging is barely affected by these attacks. Where attacks are aggressive, CARAT's advantage is large: at 40% malicious it exceeds undefended Mean by up to 47 points under ALIE and by 17–37 points under FangAttack (except CIFAR-10, α=0.1).

Presentation. Section 4 now has a short "Design rationale and evidence" paragraph. It maps each component (clipping, common-radius certificate, rank penalty, capped weights) to the failure it addresses and to the evidence above. Section 4.2 also gives the clipping-versus-normalization rationale.

Thank you for your responses. After reading the rebuttals and the authors's responses to Reviewer 2Ris, my main concerns on the design choices are not well resolved. The novelty and rationale behind the proposals are limited to me. Therefore I will maintain my score.

Reviewer_2RisOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CARAT, a class-aware aggregation method intended to protect federated learning from untargeted model-poisoning attacks when clients have non-IID data. Rather than judging client updates only by their geometric similarity or coordinate-wise statistics, CARAT tests how they affect a set of hidden, class-balanced validation tasks drawn from server-side reference data. It combines these task-level assessments with update clipping, anomaly penalties, limits on client influence, and temporal regularization.

Critical strengths
  • The paper looks beyond geometric similarity when evaluating client updates.
  • The evaluation is broad for an experimental study, and the authors report important failures.
Critical weaknesses
  • CARAT performs well in many settings, but the CIFAR-10 collapses under MinSum and MinMax are serious. Some runs reach roughly chance-level accuracy despite a malicious-client fraction of only 10%. This makes it difficult to assess the method’s practical reliability under heterogeneous data. The scope of the experiments also limits the strength of the conclusion: the main comparison uses two dataset–model combinations, full client participation, and a fixed malicious-client fraction. At higher attack fractions, the paper evaluates CARAT without comparing it against the other defenses. Those experiments therefore do not show whether CARAT maintains an advantage as the threat becomes more severe. As presented, the evidence supports competitive performance in many tested settings, but not a consistent robustness improvement.
  • CARAT combines clipping, validation-based scoring, anomaly penalties, influence constraints, and temporal regularization. This combination may be useful, but the paper should make it clearer how much the hidden-task certificate mechanism adds beyond existing reference-data-based defenses or simpler combinations of these techniques. The reference-data setup further complicates the comparison: CARAT uses 512 server reference examples, while FLTrust uses 100, and the CARAT reference examples overlap with client training data. Without a matched reference-data budget and a test using reference examples disjoint from client training data, it is hard to tell whether the observed results reflect the aggregation rule itself or differences in auxiliary data access.
Questions for authors
  • What explains the CIFAR-10 failures under MinSum and MinMax?
  • How sensitive are the results to the reference-data setup?
  • How might adaptive attackers affect CARAT’s security boundary?
3 comments

What explains the CIFAR-10 failures under MinSum and MinMax? They were caused by NaN propagation (General response item 5): a single non-finite certificate turned every aggregation weight into NaN. After the fix, all affected runs recover, and no seed is near chance level. On CIFAR-10 at 10% malicious, CARAT reaches:

αMinMaxMinSum
0.577.40 ± 0.9977.15 ± 0.71
0.155.59 ± 3.6553.63 ± 4.32

The remaining spread at α=0.1 is shared by all methods: the clean Mean reference also has a standard deviation of about 3.3 there.

Does CARAT keep its advantage as the threat becomes more severe? Yes. Figure 4 now includes FLDetector, FLTrust, MultiKrum, NormClipping and undefended Mean from 10% to 40% (General response item 1).

  • CARAT is the best defense in 22 of 24 settings.
  • Its margin grows with the threat: at 40% it is 7.6–10.1 points above the best baseline on CIFAR-100 and on CIFAR-10 with α≥0.5.
  • The only exceptions are CIFAR-10 with α=0.1 at 30–40%, the most heterogeneous setting, where FLTrust is ahead.

Sensitivity to the reference-data setup. With a pool disjoint from all client data and budgets matched at 100 and 512 examples (α=0.5, 10% malicious):

DatasetCARAT 100CARAT 512Original in-training 512
CIFAR-1076.7676.7676.85
CIFAR-10050.1650.0849.94

CARAT changes by less than 0.1 points between budgets and by at most 0.25 points from the original pool. FLTrust does not improve with 512 examples (2.6–4.2 points lower) and stays 17–22 points below CARAT. CARAT's results therefore do not rely on overlap with client data or on a larger reference budget in this setting.

What does CARAT add beyond reference-based defenses and simpler combinations?

  • Against FLTrust (single trusted direction). Under the same held-out data, CARAT is 17–22 points more accurate.
  • Against Zeno (loss reduction on server data). At 10% the two are within 0.25 points; at this fraction every defense is near the clean ceiling. CARAT additionally down-weights FangAttack, which Zeno does not (q_M 0.056–0.084 vs 0.111).
  • At high attacker fractions. At 30–40% malicious, CARAT beats all evaluated baselines in most settings, and its logs show (near-)zero weight for FangAttack in late training and for ALIE at 40% in early training. Undefended Mean collapses there (24.7% vs CARAT's 72.0% under 40% ALIE, CIFAR-10).
  • Near-mean attacks. Their influence is bounded by the cap and clipping; CARAT stays within 3 points of undefended averaging under MinMax/MinSum at every fraction.

The revision presents CARAT as an integrated rule and maps each component to the failure mode it targets (Section 4, "Design rationale and evidence").

Adaptive attackers. The hidden tasks are resampled every round from a private pool, so an attacker must improve its certificate on tasks it cannot observe. The near-mean attacks we evaluate are the evasive end of the spectrum, and their measured damage is small. Attacks that optimize directly against the certificate are an important direction, and §5 discusses them.

Thanks for the responses. I have the following concerns:

  • I am still not clear about why the previous results show lower performance than MinSum and MinMax.
  • I am not convinced by the important assumption that the server holds a set of labelled reference data to build an evaluation rule. I think this assumption may be too strong, which may reduce the contribution.
  • I have re-checked the comparison methods, and it seems that the conducted attacks and compared defensive methods are not new.

As a result, I will keep my score.

We thank the reviewer for the follow-up and address the three points in turn.

  1. Why the submitted results were lower under MinSum and MinMax on CIFAR-10.

In a few submitted runs, the hidden-task loss of the malicious candidates diverged. The scoring had in fact singled out the attackers: in every affected round, the non-finite certificates belonged exactly to the malicious clients and never to a benign one. The failure was in how the submitted implementation handled that value. A single non-finite score passed through standardization and the softmin into every aggregation weight, so the global model became NaN and accuracy dropped to chance. This was a numerical defect, not a failure of the scoring rule. The revised method treats a diverging candidate as maximally harmful (Section 4.2 and Algorithm 1), which is what the scores already indicated. With this rule, every affected run recovers, and CARAT is the strongest defense in these exact settings (CIFAR-10, 10% malicious, round-200 accuracy over five seeds):

αAttackCARATFLDetectorFLTrustMultiKrum
0.5MinMax77.4077.1660.0551.38
0.5MinSum77.1576.7862.9347.27
0.1MinMax55.5950.5743.9822.64
0.1MinSum53.6351.9145.5519.65

At α=0.1, CARAT under attack matches or exceeds the clean, unattacked Mean reference (53.53). At α=1.0 it is within 0.53 points of the best defense under both attacks.

  1. The labeled reference-data assumption.
  • A small trusted server dataset is the standard assumption of reference-based defenses such as FLTrust (NDSS 2021) and Zeno (ICML 2019). Our threat model states it explicitly (Section 3), and we do not claim CARAT is data-free.
  • With a pool held out from all client data, CARAT needs only 100 labeled examples: 10 per class on CIFAR-10 and a single example per class on CIFAR-100. With these, its accuracy is within 0.1 points of the 512-example pool and within 0.25 points of the main configuration (Appendix, reference-data control).
  • Under identical server data, CARAT is 17–22 points more accurate than FLTrust, which compresses the reference set into a single trusted direction.
  • Without semantic checks, defenses must rely on update geometry, which confuses benign non-i.i.d. diversity with poisoning. This is where CARAT gains the most: at 40% malicious clients it exceeds the strongest baseline by 7.6–10.1 points on CIFAR-100 and on CIFAR-10 with α≥0.5.

We therefore see the assumption as a modest, explicit cost that buys a large robustness gain, and the contribution as making far better use of a resource already assumed in this line of work.

  1. Choice of attacks and defenses.

We chose methods that remain the standard benchmark set in Byzantine-robust FL and that together cover each defense principle CARAT must be distinguished from:

  • distance-based selection (MultiKrum);
  • norm bounding (NormClipping);
  • a trusted reference direction (FLTrust, NDSS 2021);
  • loss on server data (Zeno);
  • history-based detection (FLDetector, KDD 2022);
  • sign statistics (SignGuard, ICDCS 2022).

The attacks include the optimization-based MinMax/MinSum (NDSS 2021), FangAttack (USENIX Security 2020), ALIE and Mimic (ICLR 2022). These remain the strongest untargeted attacks commonly used to evaluate new aggregation rules. Our goal is to test a new aggregation principle: validation on hidden, class-balanced tasks. Comparing it under identical data against both reference-based alternatives (FLTrust and Zeno) is the most direct test of that principle. If the reviewer has specific recent methods in mind, we would be glad to include them in the final version.

We hope these clarifications address the remaining concerns.

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

The paper proposes CARAT, a server-side aggregation rule for untargeted model poisoning in federated learning. From a labeled reference pool, the server samples hidden class-balanced tasks, scores robustly clipped and common-radius client updates by task-level loss reduction, and feeds those scores, together with a coordinate-rank penalty, into a capped-simplex objective with temporal regularization. On CIFAR-10/VGG19 and CIFAR-100/ResNet18, with 20 fully participating clients, five seeds, three Dirichlet regimes, and five attacks at a 10% malicious-client ratio, CARAT obtains the highest defensive accuracy in 12 of the 30 attacked settings and the second-highest in 17. The paper also reports chance-level CIFAR-10 seeds, a CARAT-only accuracy drop as the malicious fraction increases to 40%, and a breakdown of CIFAR-100 robustness at 60 clients.

Critical strengths
  1. The paper addresses a real failure mode of geometric robust aggregation: under non-i.i.d. data, a benign update can look like an outlier. Scoring candidates by loss reduction on several hidden, class-balanced server tasks is a coherent alternative to compressing the reference set into one trusted direction, and Algorithm 1 plus Tables 2–3 specify the pipeline, threat boundary, and hyperparameters.
  2. The empirical report is unusually explicit about its own limits. It uses five seeds, pairs the 12/17 ranking with the CIFAR-10 seed collapses, states that the 512-example pool is neither held out from clients nor matched to FLTrust’s 100 examples, and labels the attacker-fraction and population sweeps as CARAT-only.
Critical weaknesses
  1. The class-aware multi-task mechanism is not shown to drive the results. The only component study (Table 4) changes one factor at a time on CIFAR-100, α = 0.5, 20% malicious clients, and only ALIE and FangAttack. Full CARAT improves on T = 1, no clipping, no rank penalty, and no temporal prior by 0.13–0.80 percentage points, within seed standard deviations of 0.59–1.18. Under ALIE, one task matches eight tasks (46.68±0.93 versus 46.81±1.06). That study is also the regime in which extra tasks have the least room to differ: Table 3 uses 400 of the 512 pool examples in every CIFAR-100 task, so any two tasks share at least 288 examples. The T ablation is not run on CIFAR-10, where each task has 40 examples. The paper reports no per-class accuracy, and the attack suite is entirely untargeted and global, so class-local harm is never measured. Zeno already scores workers by validation-loss reduction on server data, but it is not a baseline.
  2. The 12/17 ranking does not isolate CARAT under a controlled comparison. The main setting gives CARAT the true attacker fraction and a 512-example training pool that remains inside client partitions, while FLTrust receives 100 examples. At 20 clients and ρ̂ = 0.1, the cap is 1/18 and Proposition A.1 forces support on at least 18 clients, so the rule can fully drop at most two clients and otherwise has little freedom to reweight; malicious weight mass is never reported. Many recorded leads are 0.16–0.54 points. On CIFAR-10, MinSum at α = 0.5 finishes at 63.58±29.96 because a seed falls to chance accuracy, and MinMax and MinSum at α = 0.1 have standard deviations of about 19–21; rankings use means only. Where a winner is named, CARAT trails FLDetector, which does not use a labeled reference set, by 0.46–1.08 points on all three CIFAR-100 Mimic settings, and it is second to FLDetector on CIFAR-10 MinSum at α = 0.5. Figure 4 and Table 5 have no baselines. In Table 5, FangAttack reaches chance-level 1% accuracy for every completed CIFAR-100 seed at 60 clients and every reported α, while unattacked CARAT accuracy rises with the population. The only runtime figure is an earlier three-seed i.i.d. diagnostic (9.94 s versus 2.80 s for Mean). Coordinate-wise median, RFA, DnC, CenteredClipping, bucketing, and SignGuard are absent. Appendix A establishes feasibility, a clipped-norm bound, order-preserving standardization, and the softmin gradient; the influence bound scales with the malicious weight q_M and does not control it.
Questions for authors
  1. If the reference pool is removed from every client partition, and FLTrust and Zeno are rerun at both 100 and 512 held-out examples, what are the round-200 accuracies, the fraction of seeds that fall to chance level, and the malicious weight q_M for CARAT and those matched baselines?
  2. On CIFAR-10, where a 40-example task can be disjoint from other draws, how do global and per-class accuracy change with T, including against an attack that directly maximizes the hidden-task scores? At 40% malicious clients and at 60 clients, do ratio-matched baselines still leave CARAT in the top two?
1 comment

We thank the reviewer for the detailed review.

Q1. Held-out pool, FLTrust and Zeno at 100 and 512 examples: accuracy, chance-level seeds, q_M.

Accuracy is in General response item 3: CARAT is 17–22 points above FLTrust under identical held-out data, and no seed of any method falls to chance level.

q_M (averaged over rounds, seeds and both budgets, which differ by at most 0.009; 10% malicious; uniform 0.10, CARAT cap 0.111):

ALIEFangAttackMinMaxMinSumMimic
CARAT (C10 / C100)0.111 / 0.1110.084 / 0.0560.111 / 0.1110.111 / 0.1110.107 / 0.105
Zeno (C10 / C100)0.111 / 0.1110.111 / 0.1110.111 / 0.1110.111 / 0.1110.107 / 0.103
FLTrust (C10 / C100)0.100 / 0.1010.008 / 0.0020.108 / 0.1030.106 / 0.1020.100 / 0.102

At 10%, CARAT down-weights FangAttack, which Zeno does not. ALIE, MinMax and MinSum are built to imitate benign statistics and keep the capped weight; their influence is bounded by the cap and clipping (Appendix A), and each lowers CARAT's accuracy by at most 1.6 points relative to its clean accuracy.

As attacks become aggressive, CARAT removes them. Per-round weights from the attacker-fraction sweep (seed 42 logs):

  • FangAttack. q_M averages at most 0.05 over the last 50 rounds in 17 of 18 settings at 20–40% malicious, and is 0 in most of them.
  • ALIE at 40%. q_M averages at most 0.06 over the first 20 rounds in all six dataset–α settings, and is exactly 0 for rounds 0–99 at α=0.5. ALIE's perturbation grows with the attacker count (z = 0.16 / 0.37 / 0.67 at 20 / 30 / 40%). Undefended Mean drops to 24.7% (CIFAR-10, α=1), while CARAT keeps 72.0%.
  • Rank penalty under FangAttack. Attackers' penalties exceed benign clients' by 2.8–17 robust (MAD) units at 30–40%.

Q2. T on CIFAR-10, per-class accuracy, a certificate-maximizing attack. These were not run and are listed as future work. On T: the CIFAR-100 ablation at 20% shows T=1 within noise of T=8, and in the reference control CARAT keeps its accuracy with one example per class. We have revised the text accordingly: the certificate is presented as a loss-based validity test within the full rule, and we do not claim that task multiplicity is independently necessary. The hidden tasks are resampled every round from a private pool, so an attacker cannot observe the tasks it is scored on.

Q3. Do ratio-matched baselines leave CARAT in the top two at 40% and at 60 clients? Yes.

  • At 40% malicious (Figure 4). CARAT is first in 5 of 6 dataset–α settings, 7.6–10.1 points ahead on CIFAR-100 and on CIFAR-10 with α≥0.5, and second in the sixth (CIFAR-10, α=0.1).
  • At 60 clients (CIFAR-100). CARAT is first for every α.
  • The 1%-accuracy FangAttack entries in the submitted Table 5 were the NaN issue and are now normal.

Other points.

  • Small leads at 10%. At 10%, even undefended Mean stays within 1–2 points of its clean accuracy, so every defense is near the ceiling. This is why we added the 20–40% comparison, where defenses differ by up to 47 points and CARAT leads clearly.
  • Mimic vs FLDetector. At 10% on CIFAR-100, CARAT is within 0.46–1.08 points of FLDetector under Mimic and leads under Mimic on CIFAR-10 at α∈{0.5, 0.1}. Mimic replays a benign update and is discussed as a limitation.
  • Support constraint at ρ̂=0.1. The cap forces support on ≥18 of 20 clients by design, a conservative trust budget that prevents any client from dominating. We now report q_M so that readers can see how the weight is used.
  • ρ̂ equals the true fraction (stated in the protocol table).
  • Additional baselines. SignGuard is now in the main comparison (General response item 4).
  • Theory. The malicious-mass bound (Appendix A) is stated in terms of q_M; the revision says so after the proposition and reports measured q_M.
  • Runtime. The labelled i.i.d. diagnostic is reported in the appendix.

Comments on the paper

1 comment

We thank the reviewers for their careful reading. The revision adds full baseline comparisons at higher attacker fractions and larger federations, a held-out and budget-matched reference-data control with Zeno, SignGuard, and the malicious weight q_M. The main findings are:

1. CARAT's advantage grows with the threat (Figure 4, 10–40% malicious, all baselines and undefended Mean, 3,600 runs).

  • CARAT is the most accurate defense in 22 of 24 settings.
  • Outside CIFAR-10 at α=0.1, its margin over the strongest baseline rises from 2.8–5.1 points at 10% to 7.6–10.1 points at 40%.
  • At 40% it exceeds undefended Mean by 5.6–7.7 points on CIFAR-100 and by 13.8 / 10.5 / 2.8 points on CIFAR-10 (α = 1 / 0.5 / 0.1). Under 40% ALIE on CIFAR-10 (α=1), Mean drops to 24.7% while CARAT keeps 72.0%.
  • The two exceptions are CIFAR-10 α=0.1 at 30% and 40%, where FLTrust is ahead (40.18 vs 35.37, 34.40 vs 33.04).

2. CARAT scales to 40 and 60 clients (client-population table, Table 5 in the submission; CIFAR-100). CARAT has the best five-attack mean in all 9 settings, 2.74–3.97 points above the strongest baseline (FLDetector). The two have similar clean accuracy, but attacks cost FLDetector 2.82–4.31 points and CARAT only 0.74–1.29.

3. CARAT's results do not depend on privileged reference data (new Appendix "Reference-Data Control, Zeno, and Malicious Weight"). We held out a fixed class-balanced pool of 512 examples, redrew client partitions on the remaining data, and gave FLTrust, Zeno and CARAT server data only from this pool, at 100 and at 512 examples (α=0.5, 10% malicious, five attacks, 300 runs):

DatasetCARAT 100CARAT 512Zeno 100Zeno 512FLTrust 100FLTrust 512
CIFAR-1076.7676.7676.7876.7559.3655.19
CIFAR-10050.1650.0850.0349.8331.7829.21
  • Under identical server data, CARAT is 17–22 points more accurate than FLTrust.
  • CARAT stays within 0.25 points of its original configuration, and 100 vs 512 examples differ by less than 0.1 points. On CIFAR-100, 100 examples means one labeled example per class, so CARAT needs very little reference data.
  • No run of any method falls to chance level. Zeno's settings (64-example batch, ρ=5×10⁻⁴, trimming as many updates as attackers) are listed in the appendix.

4. SignGuard is now in the main comparison (Figures 2–3). At 10% malicious, CARAT exceeds SignGuard under every attack and heterogeneity level where SignGuard was run (24 settings; it was not run under Mimic), by 0.26–6.22 points. The appendix also covers 20%: on the four-attack mean, CARAT is higher in all 12 dataset × α × fraction settings, by 1.22–4.73 points at 10% and 0.08–6.37 points at 20%.

5. The earlier CIFAR-10 chance-level seeds are resolved. They came from NaN propagation: one non-finite hidden-task certificate (a diverging candidate) spread through standardization and the softmin into every weight. A diverging candidate is now treated as the worst case; across all reruns this rule flagged only malicious clients, never a benign one. We reran every affected run with identical seeds. No CARAT seed in the 10% main comparison or the client-population study ends at chance level (lowest run 41.97%). In the main comparison CARAT now ranks first in 15 and second in the other 15 of 30 attacked settings (previously 12 and 17).

6. Malicious weight q_M is reported in the same appendix (see the response to jAzV).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.