acceptodds
Under review as a conference paper at ICLR 2027

History-Guided Directional Aggregation for Byzantine-Robust Federated Learning

Abstract

Defending federated learning against untargeted poisoning requires suppressing malicious influence while preserving useful information from clients with non-IID data. We propose History-Guided Directional Aggregation (HGDA), which uses a small trusted server dataset to assess clients over time and weight the distinct update directions they currently submit. At each communication round, the server trains a copy of the global model on this dataset to obtain reference updates at two checkpoints. HGDA separately accumulates positive and negative alignments between client and reference updates within a sliding window to score each client. Each current direction receives the largest score among the clients submitting it, so repeated submissions share an aggregation weight while nearby directions remain separate. The server uses the aggregated directions to update the global model with its own optimizer, independently of client and reference update magnitudes. HGDA requires no estimate of the Byzantine population. We analyze convergence with server Adam, showing how residual adversarial influence and local-training error affect progress on the benign clients' objective. Experiments on five datasets demonstrate broad Byzantine robustness across diverse poisoning attacks and Byzantine fractions under full and partial participation, with strong accuracy where representative baselines severely degrade or collapse. HGDA remains effective under poisoning when the trusted dataset is not representative of client data and has incomplete class coverage. It also demonstrates resilience to scheduled and adaptive on-off attacks, in which clients alternate between benign training and poisoning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Byzantine-Robust Federated Learning, Robust Aggregation, Model Poisoning, Data Heterogeneity

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kXgVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces History-Guided Directional Aggregation (HGDA), a server-side defense for Byzantine-robust federated learning. Each client is assigned a score derived from how well its recent update history aligns with two reference checkpoints obtained by training on a small trusted dataset. Clients whose current update directions are identical are grouped together and share a single weight determined by the highest score in the group, and the resulting aggregated direction is applied through an Adam update. The authors give a convergence analysis under stated conditions and evaluate robustness on five datasets against multiple attacks and participation settings.

Critical strengths
  • The problem is relevant: filtering malicious updates without discarding the useful contributions of heterogeneous benign clients. Pairing a trusted reference with per-client alignment histories is a sensible design choice.

  • The experiments span multiple datasets, participation regimes, and attack schedules, and the appendix contains ablations and phase-specific breakdowns that make the robustness claims easier to assess.

Critical weaknesses
  • The direction-grouping step depends on normalized updates being exactly equal. This creates an easy evasion: an attacker can take a copied benign update and add a small, distinct perturbation per malicious client, so that each copy is treated as a separate direction and receives its own weight. Each client could then keep a favorable historical score while the group as a whole regains its full aggregate influence. The Adaptive Mimic experiment does not cover this case, since all attackers still submit exactly identical updates. The paper should evaluate perturbed-copy attacks and report whether the scoring mechanism alone can bound the attackers' influence once exact grouping no longer triggers.

  • The choice of best test accuracy as the primary metric is not justified and appears to overstate the defense. In Table A.51, best accuracy reaches 85.29%, whereas the mean accuracy over attack-on rounds is only 38.00%. A gap of this size indicates that peak accuracy can hide severe degradation while the attack is active, and therefore does not faithfully reflect how well the defense protects the model.

Questions for authors
  1. How does HGDA behave when each attacker submits a copied benign update with a small, client-specific perturbation? Please sweep the perturbation magnitude and report both the attackers' combined aggregation weight and the resulting model accuracy.

  2. Do the reported gains over the baselines hold when the metric is final accuracy or average accuracy over the late training rounds, rather than best test accuracy?

9 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1)

Q1 & W1: Exact Direction Matching and Adaptive Near-Copy Evaluation

Thank you for raising this question. In our implementation, direction equality is checked by exact elementwise equality (torch.equal) on unit-normalized updates, with no angular tolerance. As motivated in Section 1, HGDA targets broad Byzantine robustness across diverse tasks, poisoning attacks, Byzantine fractions and participation conditions under non-IID data: it combines dual-checkpoint alignment histories (Eqs. 1–4, anchored in DsD_s) with unit-direction weighting (Eq. 5) that removes identical copies while preserving benign directional diversity, and we candidly noted the near-copy boundary in Section 5. To test this setting directly, we extended our adaptive Mimic attack (Section 4.4) with orthogonal perturbations across five scales: even when tiny perturbations bypass exact grouping, HGDA degrades gracefully without collapsing and outperforms the strongest Mimic baselines across the sweep.

1. Implementation and why exact matching is a reasonable design choice

Each positive-score client's nonzero update Δit\Delta_i^t is normalized to a unit vector uit=Δit/∥Δit∥2u_i^t = \Delta_i^t / \lVert \Delta_i^t \rVert_2, and updates share a group in Eq. (5) only when all coordinates match (torch.equal). Keeping Eq. (5) exact is well-motivated for three reasons:

  1. Preserving benign non-IID diversity (Section 1). Identical unit vectors add no new directional information, so assigning them one group score (max⁡i∈Gsit\max _ {i \in G} s_i^t) removes copy multiplication without losing signal. Under non-IID data, heterogeneous benign clients naturally produce nearby directions, so merging nearby updates would suppress useful benign diversity across subpopulations.
  2. Capping near-copies without harming non-IID diversity is a dedicated follow-up problem (Section 5). Because copy deduplication is one component within HGDA's broader robustness framework rather than its sole objective, we keep Eq. (5) exact and parameter-free here: simply adding an angular threshold is ineffective (an adaptive attacker can step just outside a tight threshold, while a wider one merges benign updates), and bounding collective near-copy weight while preserving benign diversity is a dedicated problem beyond this paper's scope, left for future work.
  3. Dual-reference historical scoring anchored in DsD_s acts as the safeguard. Eq. (5) handles identical copies, while distinct perturbed updates remain governed by DsD_s-anchored historical scoring (Eq. 4). Because both references are recomputed from θt\theta_t on DsD_s each round, persistent drift from DsD_s turns alignment negative and penalizes scores via accumulated opposition, so near-copies keep positive scores only while maintaining net positive windowed alignment with both references, limiting persistent pull to task-compatible benign directions rather than unchecked drift from DsD_s.

2. Adaptive near-copy Mimic across perturbation scales

Under our Section 4.1 CIFAR-10 setup (three seeds within the shared communication budget), the adaptive attacker tailors target selection to each defense's own mechanism. For HGDA, the attacker identifies the highest-scoring (most trusted) benign update Δ⋆t\Delta _ \star^t (u⋆t=Δ⋆t/∥Δ⋆t∥2u _ \star^t = \Delta _ \star^t / \lVert \Delta _ \star^t \rVert_2) each round and perturbs it along independent orthogonal unit vectors zitz_i^t while preserving ∥Δ⋆t∥2\lVert \Delta _ \star^t \rVert_2:

Δit=∥Δ⋆t∥2uit,uit=u⋆t+εzit1+ε2,∥zit∥2=1,⟨zit,u⋆t⟩=0.\Delta_i^t = \lVert \Delta _ \star^t \rVert_2 u_i^t, \quad u_i^t = \frac{u _ \star^t + \varepsilon z_i^t}{\sqrt{1 + \varepsilon^2}}, \quad \lVert z_i^t \rVert_2 = 1, \quad \langle z_i^t, u _ \star^t \rangle = 0.

Sweeping ε∈{0,10−3,0.1,1,10}\varepsilon \in \lbrace 0, 10^{-3}, 0.1, 1, 10 \rbrace across m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace% spans exact copies (ε=0\varepsilon = 0), nearly collinear copies (ε=10−3,0.1\varepsilon = 10^{-3}, 0.1 with cosines 0.9999995,0.9950.9999995, 0.995 to u⋆tu _ \star^t) that bypass torch.equal and wider perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace).

In our CIFAR-10 Mimic matrix (Table A.20), classical aggregators (Multi-Krum, RFA, Coordinate Median, Trimmed Mean and Bulyan) drop to 35.47%–57.94% or N/A at m=40m = 40% and below 40% or N/A at 80%. We therefore compare against the three external baseline families competitive at 40% under Mimic, evaluating all three at 40% and 80%: FLTrust-Early (best external baseline at 77.74% and 67.27%, beating FLTrust-Late at all fractions), Fed-NGA (best server-data-free baseline at 77.32% and 59.04%) and BOBA (68.68% at 40%). Against each baseline, the attacker likewise adapts target selection to its rule: copying the highest-trust benign update each round for FLTrust-Early, the benign client with the largest sample-count weight for Fed-NGA and a benign client near BOBA's server PCA subspace selected to minimize aggregate alignment with the server update (perturbed in that subspace).

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2)

Q1 & W1: Robustness Across Perturbation Scales and Scope

The tables below report three-seed CIFAR-10 test accuracy (%) within the shared communication budget as Last-100 mean ± SD (Best mean), where Last-100 averages the final 100 rounds, Best is each run's maximum and bold marks each column's best (Part 3 then reports the complete 20%–80% sweep pairing combined aggregation weight with model accuracy):

m=40m = 40% Byzantine clients

Methodε=0\varepsilon = 010−310^{-3}0.10.1111010
HGDA (W=100W=100)80.85±1.30 (82.19)75.93±1.04 (79.83)76.61±0.90 (80.31)78.97±0.86 (81.62)80.94±1.33 (82.35)
HGDA (W=10W=10)80.49±0.93 (82.14)73.60±1.56 (78.95)74.77±0.15 (78.84)78.13±1.60 (81.03)80.19±0.93 (81.56)
FLTrust-Early62.58±1.99 (72.59)63.35±2.51 (73.19)64.14±2.02 (73.16)68.91±1.87 (75.96)76.57±2.21 (79.84)
Fed-NGA70.82±2.10 (77.73)69.86±3.65 (77.32)70.75±3.19 (78.29)73.08±2.62 (78.37)81.27±1.17 (81.84)
BOBA59.44±1.97 (61.77)60.13±2.68 (62.00)60.27±1.98 (62.37)61.41±1.75 (63.12)59.56±2.27 (61.22)

m=80m = 80% Byzantine clients

Methodε=0\varepsilon = 010−310^{-3}0.10.1111010
HGDA (W=100W=100)71.26±0.30 (73.29)55.42±0.52 (65.81)55.15±2.03 (65.97)56.56±1.05 (66.60)70.22±0.98 (72.79)
HGDA (W=10W=10)70.86±0.89 (73.06)57.35±2.15 (64.52)55.10±2.68 (63.86)57.13±1.47 (65.94)70.45±1.13 (72.71)
FLTrust-Early50.05±0.42 (61.56)50.79±1.72 (61.43)50.93±2.00 (61.61)53.42±1.62 (63.04)66.27±1.83 (70.31)
Fed-NGA47.40±6.48 (56.61)47.39±7.88 (57.43)47.91±3.04 (56.00)51.00±0.97 (58.96)62.83±0.69 (65.96)
BOBA20.02±4.31 (25.89)20.90±5.39 (28.33)22.86±3.34 (28.20)20.41±7.54 (28.21)21.40±5.93 (26.78)

Across all five scales at m=40m = 40% and 80%, HGDA (W=100W = 100) wins 29 of 30 external-baseline comparisons on Last-100 (save Fed-NGA's 81.27±1.17% vs. 80.94±1.33% at 40%, ε=10\varepsilon = 10, within one SD) and all 30 on Best accuracy, with tighter SDs (0.30%–2.03% vs. up to 7.88%):

  1. When tiny perturbations (ε=10−3\varepsilon = 10^{-3}) bypass exact grouping, HGDA degrades gracefully at m=40m = 40%, stays bounded without collapse at m=80m = 80% and leads every baseline. At ε=0\varepsilon = 0, exact copies share one group with the copied benign client. At ε=10−3\varepsilon = 10^{-3}, near-copies split into distinct directions and receive separate weights, returning HGDA (W=100W = 100) to its client-level behavior (matching client-level HGDA under exact copies at m=80m = 80%: 55.42±0.52% vs. 55.16±1.73% Last-100, and 65.81% vs. 66.05% Best). Crucially, accuracy never collapses: at m=40m = 40%, HGDA (W=100W = 100) dips gracefully by only 4.93 Last-100 and 2.36 Best points from ε=0\varepsilon = 0, beating Fed-NGA by 6.06 Last-100 (with under one-third of its SD, 1.04% vs. 3.65%) and 2.51 Best points, while even at 4-to-1 attacker majority (m=80m = 80%) degradation remains bounded, leading FLTrust-Early by 4.63 Last-100 and 4.38 Best points.
  2. How dual-reference historical scoring (W=100W = 100), anchored in DsD_s, bounds near-copy bias (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace). Because both references are recomputed from θt\theta_t on DsD_s each round, repeatedly amplifying one skewed benign client shifts θt\theta_t toward its local classes and weakens subsequent alignment with DsD_s. A longer window (W=100W = 100 vs. 1010) further stabilizes scores against short-term fluctuations: across the full five-fraction (m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace%) by five-scale grid (see Part 3), W=100W = 100 beats W=10W = 10 in 21 of 25 Last-100 and 24 of 25 Best cells (winning 9 of 10 near-copy settings on Last-100 and all 10 of 10 on Best across 20%–80%).
  3. Larger perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace) weaken reference alignment and restore accuracy. As ε\varepsilon grows, the shared copied component u⋆tu _ \star^t shrinks by (1+ε2)−1/2(1 + \varepsilon^2)^{-1/2} while independent orthogonal components zitz_i^t average down across attackers, returning HGDA (W=100W = 100) at ε=10\varepsilon = 10 to its unperturbed ε=0\varepsilon = 0 level at both 40% and 80%.

Scope. Even when near-copies bypass Eq. (5) (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace), HGDA's DsD_s-anchored historical scoring keeps degradation bounded and outperforms every external baseline, supporting our choice to keep Eq. (5) exact here (preserving benign non-IID diversity, Section 1) while leaving dedicated near-copy weight control for future work (Section 5).

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3)

Q1 & W1: Combined Weight and Accuracy Across All 25 Settings

Following Parts 1–2, the table below pairs combined aggregation weight with model accuracy across all 25 settings (three seeds each). Unlike opposing attacks where HGDA drives attacker weight near zero, near-copy attackers submit nearly collinear copies of the highest-scoring benign update u⋆tu _ \star^t, testing how Eqs. 1–4 (DsD_s-anchored historical scoring) and Eq. 5 work together:

  • Wt (%): three-seed mean over the shared budget of total weight πGt\pi_G^t on copy-containing groups (the shared benign group at ε=0\varepsilon = 0, and the sum of distinct copy weights at ε>0\varepsilon > 0).
  • Test Acc (%): three-seed CIFAR-10 accuracy as Last-100 mean ± SD (Best mean), bolding the higher mean.
mmε\varepsilonW=100W=100 Wt (%)W=100W=100 Test Acc (%)W=10W=10 Wt (%)W=10W=10 Test Acc (%)
20%003.8583.00±0.54 (84.46)4.9582.46±0.53 (83.96)
20%10−310^{-3}30.1581.68±0.23 (83.83)33.3380.75±0.50 (83.18)
20%0.10.130.5181.62±0.55 (83.94)33.4080.75±0.42 (83.25)
20%1128.9082.68±0.11 (84.53)30.9881.82±0.71 (84.09)
20%101010.8082.63±0.62 (84.19)9.2082.00±0.12 (83.51)
40%004.5680.85±1.30 (82.19)5.6680.49±0.93 (82.14)
40%10−310^{-3}48.0175.93±1.04 (79.83)51.6073.60±1.56 (78.95)
40%0.10.148.2276.61±0.90 (80.31)51.5174.77±0.15 (78.84)
40%1145.7778.97±0.86 (81.62)47.0378.13±1.60 (81.03)
40%101021.9780.94±1.33 (82.35)19.9880.19±0.93 (81.56)
50%004.9779.44±1.00 (81.03)6.3678.72±0.89 (80.40)
50%10−310^{-3}54.5371.62±0.64 (77.64)55.7868.54±3.54 (75.80)
50%0.10.155.0071.82±1.95 (77.59)56.3267.73±1.93 (76.00)
50%1153.2976.21±0.79 (79.63)53.6674.39±0.17 (78.62)
50%101027.9779.23±1.23 (80.84)25.0278.82±1.29 (80.54)
60%006.1477.75±0.89 (79.43)7.5277.53±1.18 (79.54)
60%10−310^{-3}61.7866.37±1.82 (73.30)65.4463.84±2.03 (71.43)
60%0.10.160.9166.69±2.66 (74.40)65.1964.11±2.03 (71.80)
60%1160.3671.72±1.06 (76.94)61.5871.88±0.61 (76.82)
60%101034.6077.85±0.83 (79.29)31.4277.35±0.95 (79.04)
80%009.3171.26±0.30 (73.29)11.7370.86±0.89 (73.06)
80%10−310^{-3}79.0655.42±0.52 (65.81)83.4457.35±2.15 (64.52)
80%0.10.178.6955.15±2.03 (65.97)83.2755.10±2.68 (63.86)
80%1176.9056.56±1.05 (66.60)77.8957.13±1.47 (65.94)
80%101051.6070.22±0.98 (72.79)46.6970.45±1.13 (72.71)
  1. Exact copies (ε=0\varepsilon = 0, 20%–80%): Although copies hold 63.14%–87.69% (40%–80%) of pre-grouping client scores, Eq. (5) merges them into the copied benign client's group (3.85%–9.31% weight across 20%–80%, zero duplication gain).
  2. Near-copies (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace): graceful degradation at 20%–40% and bounded influence without collapse at 50%–80%. Once perturbations bypass torch.equal, near-copies receive client-level weights as valid benign replicas, yet W=100W = 100 lowers combined weight below W=10W = 10 in all 10/10 near-copy settings across 20%–80% (including −4.38 and −4.57 pp at 80%, and well below FLTrust-Early's 61.16%–61.22% at 40% and 87.28%–87.56% at 80%). At 20%–40%, accuracy degrades gracefully by only 1.32–4.93 Last-100 points. Even under majority corruption (50%–80%), attacker influence and accuracy remain bounded without collapse, winning 9/10 Last-100 and 10/10 Best pairs over W=10W = 10 while outperforming every baseline in Part 2 (even at 80%, leading FLTrust-Early, Fed-NGA and BOBA by +4.22–4.63, +7.24–8.03 and +32.29–34.52 Last-100 pp), because near-copies keep positive scores only while staying aligned with both DsD_s references (Part 1).
  3. Larger perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace, 20%–80%): Stronger deviations lose alignment with DsD_s, cutting weight (10.80%–51.60% across 20%–80% at ε=10\varepsilon = 10) and restoring ε=0\varepsilon = 0 accuracy (70.22% Last-100 / 72.79% Best at 80%). Together with Parts 1–2, this answers W1 directly: HGDA does not hinge on Eq. (5) alone, as DsD_s-anchored historical scoring degrades gracefully at 20%–40%, keeps near-copy influence bounded without collapse up to 80% and outperforms all baselines throughout.
Reviewer_aC9xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HGDA, a way to combine client updates in federated learning when some clients try to poison the training. The server keeps a small trusted dataset and trains a copy of the global model on it every round. This gives two reference updates, an early one and a later one. Each client update is compared with both references and the server tracks how often it agrees or disagrees over a sliding window of recent rounds. A client's score is the lower of the two reference scores. If several clients send the same direction in a round they share one weight, which is the highest score in that group. The combined direction then goes into a server side Adam optimizer. The paper also gives a convergence analysis. It links progress to how much weight the attackers still get and to the error from local training. The experiments use MNIST, CIFAR-10, FEMNIST, Purchase-100 and Tiny ImageNet. They test several poisoning attacks, different Byzantine fractions, partial participation, how well the trusted data covers the classes and on-off attacks. HGDA does well in many of these settings, especially under some attacks with a high Byzantine fraction. The main contribution is mixing history based scoring with weighting per direction.

Critical strengths
  • The paper tackles a relevant and real problem, which is keeping the useful variety between clients while limiting what attackers can do when the data is non-IID. Scoring clients over time against two trusted references is a sensible idea because it depends less on one single reference. The problem statement is well motivated since judging a client only by its current update has clear limits.
  • The paper is well organized and easy to follow. The method is explained step by step and the appendix comprises an exhaustive set of details on the experiments.
  • Putting history based scoring together with weighting per direction is useful and the ablations look at the individual design choices. Overall the experiments are exhaustive, ablations are meaningful. The tests on five datasets with several attacks, Byzantine fractions and partial participation also make the empirical side stronger.
Critical weaknesses
  • It is not clear enough how the direction grouping is done and how far it works in practice. The method only groups updates that have the same unit direction and keeps nearby directions apart. The appendix says the normalized vectors are compared for equality but it does not say if that means exact equality or some tolerance. The limitations section also admits that near copies are not grouped. So an attacker who sends copies with tiny changes may never get grouped and I did not see an experiment on this. Limiting repeated influence is the main idea of the paper, so this matters the most.
  • The link between the convergence assumptions and the broad robustness claims needs a better explanation. The bound needs the leftover attacker weight to be small enough to keep a positive descent margin under Adam's coordinate scaling and it needs the local training error to be bounded. Discussion on how these conditions connect to the attacks it tests and when the bound is actually useful is seemingly missing. Also the results are the best test accuracy within the training budget. It looks like the best round is picked using the test set itself so the numbers can be a bit optimistic. However, I would like to know if a separate validation set is used.
  • Since the method HGDA separately accumulates positive and negative alignments, is there any additional overhead added? The paper reports the buffer and computational details in the manuscript, however, it is not sufficiently analyzed. Adding a wall-clock timer, GPU utilization and FLOPs w.r.t. the baselines would help analyze the implementation complexity of the model better and hence would strengthen the complexity claims better.
Questions for authors
  • How is equality between normalized update directions checked in the algorithm? Is it exact numerical equality or a tolerance or something else? How does HGDA perform against coordinated updates that are almost the same but not exactly equal?
  • Explain what is needed for the positive descent margin in the convergence analysis? For which tested Byzantine fractions is the bound still useful? How do the leftover attacker weight and the local training error show up in the bound?
  • How is the best checkpoint chosen for the reported test accuracy? Is a separate validation set used so that the test set is only kept for the final result?
  • Report the extra server cost in time and memory from training two reference updates each round, keeping client histories and grouping directions, compared with the main baselines? I saw the big O costs in the appendix but no measured numbers. Are there any FLOPs reported?
  • How well does HGDA work when the trusted dataset is shifted in other ways than missing classes, for example strong label imbalance or a class distribution that differs from the clients? More experiments would show where the trusted reference assumption breaks.
14 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1)

Q1 & W1: Exact Direction Matching and Adaptive Near-Copy Evaluation

Thank you for raising this question. In our implementation, direction equality is checked by exact elementwise equality (torch.equal) on unit-normalized updates, with no angular tolerance. As motivated in Section 1, HGDA targets broad Byzantine robustness across diverse tasks, poisoning attacks, Byzantine fractions and participation conditions under non-IID data: it combines dual-checkpoint alignment histories (Eqs. 1–4, anchored in DsD_s) with unit-direction weighting (Eq. 5) that removes identical copies while preserving benign directional diversity, and we candidly noted the near-copy boundary in Section 5. To test this setting directly, we extended our adaptive Mimic attack (Section 4.4) with orthogonal perturbations across five scales: even when tiny perturbations bypass exact grouping, HGDA degrades gracefully without collapsing and outperforms the strongest Mimic baselines across the sweep.

1. Implementation and why exact matching is a reasonable design choice

Each positive-score client's nonzero update Δit\Delta_i^t is normalized to a unit vector uit=Δit/∥Δit∥2u_i^t = \Delta_i^t / \lVert \Delta_i^t \rVert_2, and updates share a group in Eq. (5) only when all coordinates match (torch.equal). Keeping Eq. (5) exact is well-motivated for three reasons:

  1. Preserving benign non-IID diversity (Section 1). Identical unit vectors add no new directional information, so assigning them one group score (max⁡i∈Gsit\max _ {i \in G} s_i^t) removes copy multiplication without losing signal. Under non-IID data, heterogeneous benign clients naturally produce nearby directions, so merging nearby updates would suppress useful benign diversity across subpopulations.
  2. Capping near-copies without harming non-IID diversity is a dedicated follow-up problem (Section 5). Because copy deduplication is one component within HGDA's broader robustness framework rather than its sole objective, we keep Eq. (5) exact and parameter-free here: simply adding an angular threshold is ineffective (an adaptive attacker can step just outside a tight threshold, while a wider one merges benign updates), and bounding collective near-copy weight while preserving benign diversity is a dedicated problem beyond this paper's scope, left for future work.
  3. Dual-reference historical scoring anchored in DsD_s acts as the safeguard. Eq. (5) handles identical copies, while distinct perturbed updates remain governed by DsD_s-anchored historical scoring (Eq. 4). Because both references are recomputed from θt\theta_t on DsD_s each round, persistent drift from DsD_s turns alignment negative and penalizes scores via accumulated opposition, so near-copies keep positive scores only while maintaining net positive windowed alignment with both references, limiting persistent pull to task-compatible benign directions rather than unchecked drift from DsD_s.

2. Adaptive near-copy Mimic across perturbation scales

Under our Section 4.1 CIFAR-10 setup (three seeds within the shared communication budget), the adaptive attacker tailors target selection to each defense's own mechanism. For HGDA, the attacker identifies the highest-scoring (most trusted) benign update Δ⋆t\Delta _ \star^t (u⋆t=Δ⋆t/∥Δ⋆t∥2u _ \star^t = \Delta _ \star^t / \lVert \Delta _ \star^t \rVert_2) each round and perturbs it along independent orthogonal unit vectors zitz_i^t while preserving ∥Δ⋆t∥2\lVert \Delta _ \star^t \rVert_2:

Δit=∥Δ⋆t∥2uit,uit=u⋆t+εzit1+ε2,∥zit∥2=1,⟨zit,u⋆t⟩=0.\Delta_i^t = \lVert \Delta _ \star^t \rVert_2 u_i^t, \quad u_i^t = \frac{u _ \star^t + \varepsilon z_i^t}{\sqrt{1 + \varepsilon^2}}, \quad \lVert z_i^t \rVert_2 = 1, \quad \langle z_i^t, u _ \star^t \rangle = 0.

Sweeping ε∈{0,10−3,0.1,1,10}\varepsilon \in \lbrace 0, 10^{-3}, 0.1, 1, 10 \rbrace across m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace% spans exact copies (ε=0\varepsilon = 0), nearly collinear copies (ε=10−3,0.1\varepsilon = 10^{-3}, 0.1 with cosines 0.9999995,0.9950.9999995, 0.995 to u⋆tu _ \star^t) that bypass torch.equal and wider perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace).

In our CIFAR-10 Mimic matrix (Table A.20), classical aggregators (Multi-Krum, RFA, Coordinate Median, Trimmed Mean and Bulyan) drop to 35.47%–57.94% or N/A at m=40m = 40% and below 40% or N/A at 80%. We therefore compare against the three external baseline families competitive at 40% under Mimic, evaluating all three at 40% and 80%: FLTrust-Early (best external baseline at 77.74% and 67.27%, beating FLTrust-Late at all fractions), Fed-NGA (best server-data-free baseline at 77.32% and 59.04%) and BOBA (68.68% at 40%). Against each baseline, the attacker likewise adapts target selection to its rule: copying the highest-trust benign update each round for FLTrust-Early, the benign client with the largest sample-count weight for Fed-NGA and a benign client near BOBA's server PCA subspace selected to minimize aggregate alignment with the server update (perturbed in that subspace).

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2)

Q1 & W1: Robustness Across Perturbation Scales and Scope

The tables report three-seed CIFAR-10 test accuracy (%) within the shared communication budget as Last-100 mean ± SD (Best mean), where Last-100 averages the final 100 rounds, Best is each run's maximum and bold marks each column's best:

m=40m = 40% Byzantine clients

Methodε=0\varepsilon = 010−310^{-3}0.10.1111010
HGDA (W=100W=100)80.85±1.30 (82.19)75.93±1.04 (79.83)76.61±0.90 (80.31)78.97±0.86 (81.62)80.94±1.33 (82.35)
HGDA (W=10W=10)80.49±0.93 (82.14)73.60±1.56 (78.95)74.77±0.15 (78.84)78.13±1.60 (81.03)80.19±0.93 (81.56)
FLTrust-Early62.58±1.99 (72.59)63.35±2.51 (73.19)64.14±2.02 (73.16)68.91±1.87 (75.96)76.57±2.21 (79.84)
Fed-NGA70.82±2.10 (77.73)69.86±3.65 (77.32)70.75±3.19 (78.29)73.08±2.62 (78.37)81.27±1.17 (81.84)
BOBA59.44±1.97 (61.77)60.13±2.68 (62.00)60.27±1.98 (62.37)61.41±1.75 (63.12)59.56±2.27 (61.22)

m=80m = 80% Byzantine clients

Methodε=0\varepsilon = 010−310^{-3}0.10.1111010
HGDA (W=100W=100)71.26±0.30 (73.29)55.42±0.52 (65.81)55.15±2.03 (65.97)56.56±1.05 (66.60)70.22±0.98 (72.79)
HGDA (W=10W=10)70.86±0.89 (73.06)57.35±2.15 (64.52)55.10±2.68 (63.86)57.13±1.47 (65.94)70.45±1.13 (72.71)
FLTrust-Early50.05±0.42 (61.56)50.79±1.72 (61.43)50.93±2.00 (61.61)53.42±1.62 (63.04)66.27±1.83 (70.31)
Fed-NGA47.40±6.48 (56.61)47.39±7.88 (57.43)47.91±3.04 (56.00)51.00±0.97 (58.96)62.83±0.69 (65.96)
BOBA20.02±4.31 (25.89)20.90±5.39 (28.33)22.86±3.34 (28.20)20.41±7.54 (28.21)21.40±5.93 (26.78)

Across all five scales at m=40m = 40% and 80%, HGDA (W=100W = 100) wins 29 of 30 external-baseline comparisons on Last-100 (save Fed-NGA's 81.27±1.17% vs. 80.94±1.33% at 40%, ε=10\varepsilon = 10, within one SD) and all 30 on Best accuracy, with tighter SDs (0.30%–2.03% vs. up to 7.88%):

  1. When tiny perturbations (ε=10−3\varepsilon = 10^{-3}) bypass exact grouping, HGDA degrades gracefully at m=40m = 40%, stays bounded without collapse at m=80m = 80% and leads every baseline. At ε=0\varepsilon = 0, exact copies share one group with the copied benign client. At ε=10−3\varepsilon = 10^{-3}, near-copies split into distinct directions and receive separate weights, returning HGDA (W=100W = 100) to its client-level behavior (matching client-level HGDA under exact copies at m=80m = 80%: 55.42±0.52% vs. 55.16±1.73% Last-100, and 65.81% vs. 66.05% Best). Crucially, accuracy never collapses: at m=40m = 40%, HGDA (W=100W = 100) dips gracefully by only 4.93 Last-100 and 2.36 Best points from ε=0\varepsilon = 0, beating Fed-NGA by 6.06 Last-100 (with under one-third of its SD, 1.04% vs. 3.65%) and 2.51 Best points, while even at 4-to-1 attacker majority (m=80m = 80%) degradation remains bounded, leading FLTrust-Early by 4.63 Last-100 and 4.38 Best points.
  2. How dual-reference historical scoring (W=100W = 100), anchored in DsD_s, bounds near-copy bias (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace). Because both references are recomputed from θt\theta_t on DsD_s each round, repeatedly amplifying one skewed benign client shifts θt\theta_t toward its local classes and weakens subsequent alignment with DsD_s. A longer window (W=100W = 100 vs. 1010) further stabilizes scores against short-term fluctuations: across the full five-fraction (m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace%) by five-scale grid (see Part 3), W=100W = 100 beats W=10W = 10 in 21 of 25 Last-100 and 24 of 25 Best cells (winning 9 of 10 near-copy settings on Last-100 and all 10 of 10 on Best across 20%–80%).
  3. Larger perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace) weaken reference alignment and restore accuracy. As ε\varepsilon grows, the shared copied component u⋆tu _ \star^t shrinks by (1+ε2)−1/2(1 + \varepsilon^2)^{-1/2} while independent orthogonal components zitz_i^t average down across attackers, returning HGDA (W=100W = 100) at ε=10\varepsilon = 10 to its unperturbed ε=0\varepsilon = 0 level at both 40% and 80%.

Scope. Even when near-copies bypass Eq. (5) (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace), HGDA's DsD_s-anchored historical scoring keeps degradation bounded and outperforms every external baseline, supporting our choice to keep Eq. (5) exact here (preserving benign non-IID diversity, Section 1) while leaving dedicated near-copy weight control for future work (Section 5).

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3)

Q1 & W1: Complete 20%–80% Near-Copy Grid for HGDA (W=100W = 100 vs. W=10W = 10)

Following Part 2, the table below reports the full five-fraction (m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace%) by five-scale (ε∈{0,10−3,0.1,1,10}\varepsilon \in \lbrace 0, 10^{-3}, 0.1, 1, 10 \rbrace) adaptive near-copy Mimic grid comparing HGDA (W=100W = 100) and HGDA (W=10W = 10). Each cell reports three-seed CIFAR-10 test accuracy (%) within the shared communication budget as Last-100 mean ± SD (Best mean), and bold marks the higher mean between W=100W = 100 and W=10W = 10 at each (m,ε)(m, \varepsilon) pair:

mmWindowε=0\varepsilon = 010−310^{-3}0.10.1111010
20%W=100W = 10083.00±0.54 (84.46)81.68±0.23 (83.83)81.62±0.55 (83.94)82.68±0.11 (84.53)82.63±0.62 (84.19)
20%W=10W = 1082.46±0.53 (83.96)80.75±0.50 (83.18)80.75±0.42 (83.25)81.82±0.71 (84.09)82.00±0.12 (83.51)
40%W=100W = 10080.85±1.30 (82.19)75.93±1.04 (79.83)76.61±0.90 (80.31)78.97±0.86 (81.62)80.94±1.33 (82.35)
40%W=10W = 1080.49±0.93 (82.14)73.60±1.56 (78.95)74.77±0.15 (78.84)78.13±1.60 (81.03)80.19±0.93 (81.56)
50%W=100W = 10079.44±1.00 (81.03)71.62±0.64 (77.64)71.82±1.95 (77.59)76.21±0.79 (79.63)79.23±1.23 (80.84)
50%W=10W = 1078.72±0.89 (80.40)68.54±3.54 (75.80)67.73±1.93 (76.00)74.39±0.17 (78.62)78.82±1.29 (80.54)
60%W=100W = 10077.75±0.89 (79.43)66.37±1.82 (73.30)66.69±2.66 (74.40)71.72±1.06 (76.94)77.85±0.83 (79.29)
60%W=10W = 1077.53±1.18 (79.54)63.84±2.03 (71.43)64.11±2.03 (71.80)71.88±0.61 (76.82)77.35±0.95 (79.04)
80%W=100W = 10071.26±0.30 (73.29)55.42±0.52 (65.81)55.15±2.03 (65.97)56.56±1.05 (66.60)70.22±0.98 (72.79)
80%W=10W = 1070.86±0.89 (73.06)57.35±2.15 (64.52)55.10±2.68 (63.86)57.13±1.47 (65.94)70.45±1.13 (72.71)

Across all 25 settings (20%–80%), W=100W = 100 outperforms W=10W = 10 in 21 of 25 Last-100 and 24 of 25 Best comparisons. In the near-copy regime (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace) where exact grouping no longer merges copies, HGDA (W=100W = 100) degrades gracefully at 20%–40% (only 1.32–4.93 Last-100 points from ε=0\varepsilon = 0) and keeps degradation bounded without collapse across majority corruption (50%–80%), winning 9 of 10 Last-100 and all 10 of 10 Best comparisons over W=10W = 10 across 20%–80% (including +3.08+3.08 and +4.09+4.09 points at 50%, +2.53+2.53 and +2.58+2.58 points at 60% and higher Best at 80%) while outperforming every external baseline in Part 2.

Reviewer_eLU7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HGDA, a robust aggregation method for federated learning under untargeted Byzantine attacks. The method utilizes a small trusted server dataset to obtain two reference directions at different training checkpoints. Clients are scored based on their recent positive and negative cosine alignments with these references. The server then aggregates normalized client updates, assigning a single weight to each distinct direction based on the highest client score within that group. Finally, Server Adam determines the global model update.

Critical strengths
  • The paper makes a sensible distinction between assessing client reliability over time and determining the weight of their current update directions. The work provides a direct test of the direction-grouping rule: under adaptive Mimic attacks, replacing client-level weighting with direction-level weighting improves CIFAR-10 accuracy from 64.14% to 71.83% (Table 2(c)), while keeping the scoring rule and server optimizer fixed. This offers strong evidence that the grouping mechanism effectively mitigates the exact-copy attacks it is designed to address.

  • The experimental evaluation is comprehensive, covering full and partial participation, high Byzantine fractions, and various poisoning types. The experiments involving incomplete trusted-set class coverage are particularly relevant, as they demonstrate that HGDA can still leverage client data rather than relying solely on the limited server dataset. Additionally, the paper is explicit about baseline feasibility constraints and includes results on changing attacker behavior, making the empirical findings more informative than a simple comparison on standard attack settings.

Critical weaknesses
  • The protection against repeated influence appears limited. In Eq. (5), two updates are grouped only if their normalized directions are identical. A small perturbation can fragment a single group into multiple groups while leaving the cosine alignments, and potentially the client scores, nearly unchanged. Consequently, coordinated attackers could recover multiple weights for essentially the same direction. Although the paper acknowledges nearby copies as a limitation, this deserves an explicit attack experiment, given that direction-level aggregation is a core contribution. Furthermore, there is an unresolved question regarding model buffers. Appendix A.2 states that BN statistics are averaged using client sample counts, apparently outside the proposed scoring and aggregation rules. It is unclear whether an attacker can exploit this separate update path. Both points significantly affect the scope of the robustness claim.

  • The reliance on best test accuracy for evaluating robustness is concerning. The main results report the highest accuracy achieved anywhere within the training budget, which obscures the model's reliability during active attacks. Table A.51 illustrates this issue sharply: with the (1,2) label-flip schedule and W=100, HGDA reaches a best accuracy of 85.29%, but averages only 38.00% during attack-on rounds, compared to 83.81% during attack-off rounds. While the method can recover, this is distinct from preventing severe degradation during the attack phase. This distinction is not sufficiently reflected in the main evaluation. Additionally, drawing strong conclusions from the CIFAR-10 grid is difficult, as most configurations rely on a single seed.

Questions for authors
  1. Has HGDA been tested against approximate Mimic attacks, where malicious clients submit slightly perturbed versions of the same update? How exactly is direction equality checked in the implementation?

  2. Are BN running statistics aggregated from all participating clients, including those with zero HGDA scores? Can Byzantine clients modify these buffers independently of their parameter updates? If so, what mechanisms prevent manipulation of the model through this path?

  3. Could u report final accuracy and attack-on accuracy for the main on-off experiments, rather than relying primarily on the best checkpoint? Specifically, how should the large accuracy drop during the attack phase in Table A.51 be interpreted when assessing the method's resilience?

  4. Are additional independent runs available for the more challenging CIFAR-10 settings? Also, were the Purchase-100 learning-rate tuning runs, which included attacked configurations, separate from the final evaluation? How were the baseline hyperparameters selected?

16 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1)

Q1 & W1: Exact Direction Matching and Adaptive Near-Copy Evaluation

Thank you for raising this question. In our implementation, direction equality is checked by exact elementwise equality (torch.equal) on unit-normalized updates, with no angular tolerance. As motivated in Section 1, HGDA targets broad Byzantine robustness across diverse tasks, poisoning attacks, Byzantine fractions and participation conditions under non-IID data: it combines dual-checkpoint alignment histories (Eqs. 1–4, anchored in DsD_s) with unit-direction weighting (Eq. 5) that removes identical copies while preserving benign directional diversity, and we candidly noted the near-copy boundary in Section 5. To test this setting directly, we extended our adaptive Mimic attack (Section 4.4) with orthogonal perturbations across five scales: even when tiny perturbations bypass exact grouping, HGDA degrades gracefully without collapsing and outperforms the strongest Mimic baselines across the sweep.

1. Implementation and why exact matching is a reasonable design choice

Each positive-score client's nonzero update Δit\Delta_i^t is normalized to a unit vector uit=Δit/∥Δit∥2u_i^t = \Delta_i^t / \lVert \Delta_i^t \rVert_2, and updates share a group in Eq. (5) only when all coordinates match (torch.equal). Keeping Eq. (5) exact is well-motivated for three reasons:

  1. Preserving benign non-IID diversity (Section 1). Identical unit vectors add no new directional information, so assigning them one group score (max⁡i∈Gsit\max _ {i \in G} s_i^t) removes copy multiplication without losing signal. Under non-IID data, heterogeneous benign clients naturally produce nearby directions, so merging nearby updates would suppress useful benign diversity across subpopulations.
  2. Capping near-copies without harming non-IID diversity is a dedicated follow-up problem (Section 5). Because copy deduplication is one component within HGDA's broader robustness framework rather than its sole objective, we keep Eq. (5) exact and parameter-free here: simply adding an angular threshold is ineffective (an adaptive attacker can step just outside a tight threshold, while a wider one merges benign updates), and bounding collective near-copy weight while preserving benign diversity is a dedicated problem beyond this paper's scope, left for future work.
  3. Dual-reference historical scoring anchored in DsD_s acts as the safeguard. Eq. (5) handles identical copies, while distinct perturbed updates remain governed by DsD_s-anchored historical scoring (Eq. 4). Because both references are recomputed from θt\theta_t on DsD_s each round, persistent drift from DsD_s turns alignment negative and penalizes scores via accumulated opposition, so near-copies keep positive scores only while maintaining net positive windowed alignment with both references, limiting persistent pull to task-compatible benign directions rather than unchecked drift from DsD_s.

2. Adaptive near-copy Mimic across perturbation scales

Under our Section 4.1 CIFAR-10 setup (three seeds within the shared communication budget), the adaptive attacker tailors target selection to each defense's own mechanism. For HGDA, the attacker identifies the highest-scoring (most trusted) benign update Δ⋆t\Delta _ \star^t (u⋆t=Δ⋆t/∥Δ⋆t∥2u _ \star^t = \Delta _ \star^t / \lVert \Delta _ \star^t \rVert_2) each round and perturbs it along independent orthogonal unit vectors zitz_i^t while preserving ∥Δ⋆t∥2\lVert \Delta _ \star^t \rVert_2:

Δit=∥Δ⋆t∥2uit,uit=u⋆t+εzit1+ε2,∥zit∥2=1,⟨zit,u⋆t⟩=0.\Delta_i^t = \lVert \Delta _ \star^t \rVert_2 u_i^t, \quad u_i^t = \frac{u _ \star^t + \varepsilon z_i^t}{\sqrt{1 + \varepsilon^2}}, \quad \lVert z_i^t \rVert_2 = 1, \quad \langle z_i^t, u _ \star^t \rangle = 0.

Sweeping ε∈{0,10−3,0.1,1,10}\varepsilon \in \lbrace 0, 10^{-3}, 0.1, 1, 10 \rbrace across m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace% spans exact copies (ε=0\varepsilon = 0), nearly collinear copies (ε=10−3,0.1\varepsilon = 10^{-3}, 0.1 with cosines 0.9999995,0.9950.9999995, 0.995 to u⋆tu _ \star^t) that bypass torch.equal and wider perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace).

In our CIFAR-10 Mimic matrix (Table A.20), classical aggregators (Multi-Krum, RFA, Coordinate Median, Trimmed Mean and Bulyan) drop to 35.47%–57.94% or N/A at m=40m = 40% and below 40% or N/A at 80%. We therefore compare against the three external baseline families competitive at 40% under Mimic, evaluating all three at 40% and 80%: FLTrust-Early (best external baseline at 77.74% and 67.27%, beating FLTrust-Late at all fractions), Fed-NGA (best server-data-free baseline at 77.32% and 59.04%) and BOBA (68.68% at 40%). Against each baseline, the attacker likewise adapts target selection to its rule: copying the highest-trust benign update each round for FLTrust-Early, the benign client with the largest sample-count weight for Fed-NGA and a benign client near BOBA's server PCA subspace selected to minimize aggregate alignment with the server update (perturbed in that subspace).

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2)

Q1 & W1: Robustness Across Perturbation Scales and Scope

The tables report three-seed CIFAR-10 test accuracy (%) within the shared communication budget as Last-100 mean ± SD (Best mean), where Last-100 averages the final 100 rounds, Best is each run's maximum and bold marks each column's best:

m=40m = 40% Byzantine clients

Methodε=0\varepsilon = 010−310^{-3}0.10.1111010
HGDA (W=100W=100)80.85±1.30 (82.19)75.93±1.04 (79.83)76.61±0.90 (80.31)78.97±0.86 (81.62)80.94±1.33 (82.35)
HGDA (W=10W=10)80.49±0.93 (82.14)73.60±1.56 (78.95)74.77±0.15 (78.84)78.13±1.60 (81.03)80.19±0.93 (81.56)
FLTrust-Early62.58±1.99 (72.59)63.35±2.51 (73.19)64.14±2.02 (73.16)68.91±1.87 (75.96)76.57±2.21 (79.84)
Fed-NGA70.82±2.10 (77.73)69.86±3.65 (77.32)70.75±3.19 (78.29)73.08±2.62 (78.37)81.27±1.17 (81.84)
BOBA59.44±1.97 (61.77)60.13±2.68 (62.00)60.27±1.98 (62.37)61.41±1.75 (63.12)59.56±2.27 (61.22)

m=80m = 80% Byzantine clients

Methodε=0\varepsilon = 010−310^{-3}0.10.1111010
HGDA (W=100W=100)71.26±0.30 (73.29)55.42±0.52 (65.81)55.15±2.03 (65.97)56.56±1.05 (66.60)70.22±0.98 (72.79)
HGDA (W=10W=10)70.86±0.89 (73.06)57.35±2.15 (64.52)55.10±2.68 (63.86)57.13±1.47 (65.94)70.45±1.13 (72.71)
FLTrust-Early50.05±0.42 (61.56)50.79±1.72 (61.43)50.93±2.00 (61.61)53.42±1.62 (63.04)66.27±1.83 (70.31)
Fed-NGA47.40±6.48 (56.61)47.39±7.88 (57.43)47.91±3.04 (56.00)51.00±0.97 (58.96)62.83±0.69 (65.96)
BOBA20.02±4.31 (25.89)20.90±5.39 (28.33)22.86±3.34 (28.20)20.41±7.54 (28.21)21.40±5.93 (26.78)

Across all five scales at m=40m = 40% and 80%, HGDA (W=100W = 100) wins 29 of 30 external-baseline comparisons on Last-100 (save Fed-NGA's 81.27±1.17% vs. 80.94±1.33% at 40%, ε=10\varepsilon = 10, within one SD) and all 30 on Best accuracy, with tighter SDs (0.30%–2.03% vs. up to 7.88%):

  1. When tiny perturbations (ε=10−3\varepsilon = 10^{-3}) bypass exact grouping, HGDA degrades gracefully at m=40m = 40%, stays bounded without collapse at m=80m = 80% and leads every baseline. At ε=0\varepsilon = 0, exact copies share one group with the copied benign client. At ε=10−3\varepsilon = 10^{-3}, near-copies split into distinct directions and receive separate weights, returning HGDA (W=100W = 100) to its client-level behavior (matching client-level HGDA under exact copies at m=80m = 80%: 55.42±0.52% vs. 55.16±1.73% Last-100, and 65.81% vs. 66.05% Best). Crucially, accuracy never collapses: at m=40m = 40%, HGDA (W=100W = 100) dips gracefully by only 4.93 Last-100 and 2.36 Best points from ε=0\varepsilon = 0, beating Fed-NGA by 6.06 Last-100 (with under one-third of its SD, 1.04% vs. 3.65%) and 2.51 Best points, while even at 4-to-1 attacker majority (m=80m = 80%) degradation remains bounded, leading FLTrust-Early by 4.63 Last-100 and 4.38 Best points.
  2. How dual-reference historical scoring (W=100W = 100), anchored in DsD_s, bounds near-copy bias (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace). Because both references are recomputed from θt\theta_t on DsD_s each round, repeatedly amplifying one skewed benign client shifts θt\theta_t toward its local classes and weakens subsequent alignment with DsD_s. A longer window (W=100W = 100 vs. 1010) further stabilizes scores against short-term fluctuations: across the full five-fraction (m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace%) by five-scale grid (see Part 3), W=100W = 100 beats W=10W = 10 in 21 of 25 Last-100 and 24 of 25 Best cells (winning 9 of 10 near-copy settings on Last-100 and all 10 of 10 on Best across 20%–80%).
  3. Larger perturbations (ε∈{1,10}\varepsilon \in \lbrace 1, 10 \rbrace) weaken reference alignment and restore accuracy. As ε\varepsilon grows, the shared copied component u⋆tu _ \star^t shrinks by (1+ε2)−1/2(1 + \varepsilon^2)^{-1/2} while independent orthogonal components zitz_i^t average down across attackers, returning HGDA (W=100W = 100) at ε=10\varepsilon = 10 to its unperturbed ε=0\varepsilon = 0 level at both 40% and 80%.

Scope. Even when near-copies bypass Eq. (5) (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace), HGDA's DsD_s-anchored historical scoring keeps degradation bounded and outperforms every external baseline, supporting our choice to keep Eq. (5) exact here (preserving benign non-IID diversity, Section 1) while leaving dedicated near-copy weight control for future work (Section 5).

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3)

Q1 & W1: Complete 20%–80% Near-Copy Grid for HGDA (W=100W = 100 vs. W=10W = 10)

Following Part 2, the table below reports the full five-fraction (m∈{20,40,50,60,80}m \in \lbrace 20, 40, 50, 60, 80 \rbrace%) by five-scale (ε∈{0,10−3,0.1,1,10}\varepsilon \in \lbrace 0, 10^{-3}, 0.1, 1, 10 \rbrace) adaptive near-copy Mimic grid comparing HGDA (W=100W = 100) and HGDA (W=10W = 10). Each cell reports three-seed CIFAR-10 test accuracy (%) within the shared communication budget as Last-100 mean ± SD (Best mean), and bold marks the higher mean between W=100W = 100 and W=10W = 10 at each (m,ε)(m, \varepsilon) pair:

mmWindowε=0\varepsilon = 010−310^{-3}0.10.1111010
20%W=100W = 10083.00±0.54 (84.46)81.68±0.23 (83.83)81.62±0.55 (83.94)82.68±0.11 (84.53)82.63±0.62 (84.19)
20%W=10W = 1082.46±0.53 (83.96)80.75±0.50 (83.18)80.75±0.42 (83.25)81.82±0.71 (84.09)82.00±0.12 (83.51)
40%W=100W = 10080.85±1.30 (82.19)75.93±1.04 (79.83)76.61±0.90 (80.31)78.97±0.86 (81.62)80.94±1.33 (82.35)
40%W=10W = 1080.49±0.93 (82.14)73.60±1.56 (78.95)74.77±0.15 (78.84)78.13±1.60 (81.03)80.19±0.93 (81.56)
50%W=100W = 10079.44±1.00 (81.03)71.62±0.64 (77.64)71.82±1.95 (77.59)76.21±0.79 (79.63)79.23±1.23 (80.84)
50%W=10W = 1078.72±0.89 (80.40)68.54±3.54 (75.80)67.73±1.93 (76.00)74.39±0.17 (78.62)78.82±1.29 (80.54)
60%W=100W = 10077.75±0.89 (79.43)66.37±1.82 (73.30)66.69±2.66 (74.40)71.72±1.06 (76.94)77.85±0.83 (79.29)
60%W=10W = 1077.53±1.18 (79.54)63.84±2.03 (71.43)64.11±2.03 (71.80)71.88±0.61 (76.82)77.35±0.95 (79.04)
80%W=100W = 10071.26±0.30 (73.29)55.42±0.52 (65.81)55.15±2.03 (65.97)56.56±1.05 (66.60)70.22±0.98 (72.79)
80%W=10W = 1070.86±0.89 (73.06)57.35±2.15 (64.52)55.10±2.68 (63.86)57.13±1.47 (65.94)70.45±1.13 (72.71)

Across all 25 settings (20%–80%), W=100W = 100 outperforms W=10W = 10 in 21 of 25 Last-100 and 24 of 25 Best comparisons. In the near-copy regime (ε∈{10−3,0.1}\varepsilon \in \lbrace 10^{-3}, 0.1 \rbrace) where exact grouping no longer merges copies, HGDA (W=100W = 100) degrades gracefully at 20%–40% (only 1.32–4.93 Last-100 points from ε=0\varepsilon = 0) and keeps degradation bounded without collapse across majority corruption (50%–80%), winning 9 of 10 Last-100 and all 10 of 10 Best comparisons over W=10W = 10 across 20%–80% (including +3.08+3.08 and +4.09+4.09 points at 50%, +2.53+2.53 and +2.58+2.58 points at 60% and higher Best at 80%) while outperforming every external baseline in Part 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.