acceptodds
Under review as a conference paper at ICLR 2027

Granularity Blind Spots in Byzantine-Robust Federated Learning: Shapley-Guided Subspace Poisoning of Whole-Update Defenses

Abstract

Federated learning (FL) exposes its optimization process to model poisoning by compromised clients. Byzantine-robust aggregators inspect complete updates using distances, coordinate statistics, or alignment with a reference. Because of the resulting mismatch in granularity, non-IID honest updates may appear anomalous, whereas malicious changes in a compact parameter subspace may evade detection. We demonstrate the systematic exploitability of this mismatch by introducing GRASP, a wrapper that identifies influential parameter blocks for each base attack and concentrates malicious perturbations in the corresponding subspace. GRASP ranks blocks in the target model using a scalable second-order Shapley approximation. This approximation captures individual contributions and cooperative effects within each layer without enumerating coalitions. It then injects the selected residual into the reference update and chooses the largest scale within an adaptive Euclidean radius. Across five poisoning strategies and nine Byzantine-robust defenses on MNIST, Fashion-MNIST, and CIFAR-10, GRASP exceeds the base attack in 85 of 95 settings (89.5%). It raises the average Attack Impact from 22.17 to 33.02 points. Under FLAME at a 40% malicious client fraction, it raises the malicious admission rate from 28.09% to 99.80%. These results expose a systematic blind spot in defenses that assess complete updates and motivate robustness mechanisms that operate at a finer granularity.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Model Poisoning, Shapley Values, Structured Subspace Poisoning, Byzantine-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 2 reviews, above 27% of the others.
Reviewer_qxukOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper offers an attack enhancement method for Byzantine-robust federated learning, called GASP. The method starts with the assumption that the number of attackers is less than half of the clients and the attacker is omnicent, attacker knows all the client updates. The idea is to first find the crucial weights of the target model in each layer based on the loss and compare the updates with the reference set of generated anchor updates, and inject the attack residuals into the selected weights with a feasible scale controlled by a constant \lambda. The paper monitors the increase in the attack success rate and malicious admission rate of 5 different attacks against 9 different defenses and shows an increase in the attack success rate. The paper shows strengths in the aspects of experimental improvement, attack-agnostic methodology, and organized narration. However, it shows weaknesses in the aspects of the omniscient assumption and reference set-dependent methodology, model structure, and task simplicity.

Critical strengths
  • The paper is well written and organized, with easy-to-follow sections that make the idea easy to grasp.
  • Attack-agnostic enhancement makes the methodology generalizable and plug-and-play.
  • The experiment clarity and improvement in the attack success rate by the GRASP.
Critical weaknesses
  • The method uses simple tasks such as MNIST and CIFAR and performs on CNN-like architectures. These are still credible datasets and architectures, but I wonder how this strategy works in LLM/VLM based architectures like more focused on foundational models. This will raise the impact of the methodology.
  • The credibility of the omniscient adversary is questionable; even though the authors provided a strategy that omits this, it still uses the compromised clients' pre-attack updates.
  • The method could be highly dependent on the reference set, what if there is significant difference between this set and the clients' sets.
  • As a comparison only, Pill (Guo et. al, 2026) is given as another enhancement method.
Questions for authors
  • Does the method distinguish the honest non-IID update from a malicious update?
  • The method is only attack oriented, but can this method be reversed to create a defense mechanism against the attacks or also improve the defense methods?
  • For n=3 and m=1, we have q=1 which means there is no candidate model set to select majority. In this case, how does the GRASP behave?
  • Why do you treat each layer with equal importance? Maybe some layers are more important than others.
6 comments
AuthorsOpenReview ↗

Response to Reviewer qxuk (1/4): Summary of New Results; W1 – Foundation Models and Architectures

We thank the reviewer for the careful reading and for the constructive questions. We address the four weaknesses and four questions below. Revisions are marked in green; the new diagnostic in W3 is in the anonymous repository.

Throughout, Attack Impact (AI) denotes clean accuracy minus attacked accuracy, in percentage points (pp); a pre-attack update is the update a compromised client obtains by training on its own data in the current round, i.e., what it would submit before the adversary replaces it with the malicious update.

Summary of new results

  • Attacker-local variant. Using only compromised clients' pre-attack updates, GRASP raises mean AI from 22.13 to 34.85 over the 33 fully attacker-local settings, essentially matching default GRASP (34.89); over all 66 settings the variant averages 34.17 vs. 34.83 for default (W2).
  • Reference-set mismatch. Under 90% two-class skew of the local references, the calibrated radius error is −1.1% and the anchor stays at 0.36–0.40 τH\tau_H (W3).
  • Comparison with Pill. Pill is the only directly comparable enhancement wrapper; GRASP exceeds it in 176/198 settings (W4).
  • Foundation models. GRASP's block partition maps directly to Transformer and LoRA weights, so the method extends without architectural change (W1).
  • Boundary case. GRASP is well defined at n=3n=3, m=1m=1: q=1q=1 counts honest reference supporters, not the majority size (Q3).
  • Layer budget. A per-layer quota outperforms a global cross-layer ranking by 10.36 AI on LASA (Q4).

W1. Foundation models and architectures

How does GRASP work on LLM/VLM-style architectures?

Scope first. We do not report end-to-end foundation-model experiments in this paper. Two settings are outside the current formulation and would need a different scoring objective: (i) protocols that aggregate the low-rank product rather than the LoRA factors, i.e., the product BABA of the adapter ΔW=BA\Delta W=BA, and (ii) generative tasks, where "clean accuracy minus attacked accuracy" is not the natural target. What we can establish here is that the method is not tied to CNNs and maps to Transformer-style architectures without changing its formulation; we give that mapping below.

Evaluation scope so far. Our evaluation follows the setting in which the base attacks and defenses we wrap were proposed and benchmarked, from LeNet on MNIST to ResNet-18 on CIFAR-10 (11.17M parameters, 972 blocks). Extending GRASP to federated adaptation of foundation models is a natural next step.

GRASP does not depend on the architecture. It requires only (i) a block partition of the trainable parameters, (ii) per-example gradients on a small attacker-held scoring panel, and (iii) Euclidean geometry over flattened updates. Each has a direct counterpart in Transformers:

  • Blocks. GRASP's blocks are groups of weight rows or output channels. In Transformers these are rows of the attention and MLP projections, or rows of the LoRA factors when only adapters are trained.
  • Adapters. In common federated LoRA protocols clients upload the adapter factors and the server averages them. GRASP then operates on the adapter updates, i.e., exactly the space the defense inspects, and the injection ray z(λ)=a+λδz(\lambda)=a+\lambda\delta and its distance constraints remain linear in λ\lambda.
  • Cost. Per-example gradients and block scores are needed only for the PP trainable parameters; LoRA adapters are typically comparable to or smaller than the ResNet-18 we already evaluate. The backbone forward and backward passes are the same ones each client already runs in local training.
  • Objective. For discriminative adaptation, such as ViT image classification or LLM text classification, the classification loss and Attack Impact carry over directly.

An end-to-end evaluation on foundation models is a natural next step; the two settings noted above each require a dedicated formulation rather than a re-parameterization of the current one, which we leave to future work.

AuthorsOpenReview ↗

Response to Reviewer qxuk (2/4): W2 – Attacker-Local Variant; W3 – Dependence on the Reference Set

W2. The omniscient adversary and the attacker-local variant

The attacker-local variant still uses compromised clients' pre-attack updates.

We clarify what this requirement involves and what the variant demonstrates.

What a pre-attack update is. In round tt, a compromised client trains on its own data exactly as an honest client would, producing update uju_j. This is what it would have submitted before the adversary replaces it with the malicious one; it is neither an honest client's update nor a history collected before the attack begins. Any adversary that controls a client's data and training, as our threat model assumes, can compute it. For example, in cross-device FL an adversary can register or compromise several devices as Sybil clients (Fung et al., 2020) and run standard local training on each; the pre-attack update is simply the output of that training. Prior colluding attacks rely on the same information in their partial-knowledge settings: ALIE estimates coordinate statistics from the compromised workers' own updates (Baruch et al., 2019), and Fang et al. (2020) use the before-attack local models on compromised devices. An adversary without this information could not run local training.

Evidence. Appendix D, Table 15 evaluates GRASP with an attacker-local reference set built only from these updates. Over the 33 fully attacker-local settings (Sign Flipping, ALIE, and PoisonedFL × 11 defenses), the variant raises mean Attack Impact from 22.13 to 34.85 and exceeds Base in 28 of them; default omniscient GRASP reaches 34.89 on the same settings. Across all 66 settings (six attacks × eleven defenses), the variant averages 34.17, versus 34.83 for default GRASP and 24.35 for Base.

Scope. IPM, Min-Max, and Min-Sum still use honest updates through their base attacks, and the variant is evaluated on the MNIST configuration. We will add a concise parenthetical definition of pre-attack updates to the Threat Model in the revision.


W3. Dependence on the reference set

What happens if the reference set differs substantially from the clients' sets?

Where a mismatch can arise. RR is a set of current-round client updates, not a separate dataset: GRASP takes the anchor from RR and the radius τ\tau from its pairwise distances. Under the default threat model, RR is exactly the set of honest updates, so no mismatch exists. A mismatch can arise only in the attacker-local variant, where RR consists of the compromised clients' pre-attack updates (W2).

How the variant sets the radius. The local set is smaller than the honest population (here 10 vs. 40 updates), so a percentile of its pairwise distances underestimates the honest radius τH\tau_H. The variant rescales it by γB=max⁡{1, b(n−m)/b(∣R∣)}\gamma_B=\max\{1,\,b(n-m)/b(|R|)\}, where b(k)=Φ−1 ⁣((k−3/8)/(k+1/4))b(k)=\Phi^{-1}\!\left((k-3/8)/(k+1/4)\right) is Blom's approximation to the expected maximum of kk standard normals (Appendix D). The factor depends only on the set sizes (γB≈1.39\gamma_B\approx1.39 here) and uses no honest update.

Existing evidence. All attacker-local experiments use Dirichlet α=0.1\alpha=0.1, so the compromised clients' data already differ substantially from the honest clients'. The variant still matches default GRASP end to end (mean AI 34.17 vs. 34.83 over all 66 settings; Table 15).

New diagnostic. To test a stronger mismatch, we rebuilt the 10 local references from batches in which about 60% or 90% of the examples come from two classes, on normal MNIST training trajectories (5 seeds, 10 checkpoints each), with the honest updates, model state, and target update fixed. We measure (i) the signed error of the local radius relative to τH\tau_H, which bounds the injection scale, and (ii) the distance from the local anchor to its nearest honest update, divided by τH\tau_H, which shows whether the anchor lies within the honest population.

As the diagnostic table shows, with the original references (α=0.1\alpha=0.1) the uncalibrated radius error is −23.3% and the calibrated error is +6.9%. Under 90% two-class skew the uncalibrated error grows only to −29.1%, and the calibrated error is −1.1% (per seed −12.3% to +10.3%). The anchor distance is 0.36–0.42 τH\tau_H with the original references and 0.36–0.40 τH\tau_H under 90% skew. Most of the radius error therefore comes from the smaller set size, which calibration corrects; class skew adds about 6 points of error and barely moves the anchor.

AuthorsOpenReview ↗

Response to Reviewer qxuk (3/4): W4 – Pill as Baseline; Q1 – Honest Non-IID vs. Malicious; Q2 – Use as a Defense

W4. Pill as the only enhancement baseline

If we understand correctly, the concern is that GRASP is positioned against Pill as the only enhancement baseline. We clarify the comparison on two levels.

Scope of comparable methods. To the best of our knowledge, Pill (Guo et al., 2026) is the only directly comparable enhancement method we are aware of: an attack-agnostic wrapper that strengthens an arbitrary base attack by confining it to a parameter subspace. The other methods in our evaluation (Sign Flipping, ALIE, IPM, Min-Max, Min-Sum) are base attacks that GRASP wraps, not competing wrappers, so they are the objects of enhancement rather than baselines for it. Mimic serves as an attack reference in our ablations, not as a second enhancer.

Contribution over Pill. GRASP is not merely Pill with its heuristics swapped. Across all 198 FedSGD attack–defense settings (Tables 2, 4, and 5), GRASP exceeds Pill in 176/198 (88.9%) and the base attack in 180/198 (90.9%). The controlled ablations isolate the new components: the full method reaches an average Attack Impact of 42.69, versus 32.30 for the Pill-style scaling ablation. Please refer to our response to Reviewer NHGR (Weakness 4) for a detailed mapping of each design choice to the problem it addresses relative to Pill.


Q1. Does the method distinguish an honest non-IID update from a malicious update?

GRASP does not need to make this distinction, because its reference set is not sampled from the submitted updates.

How RR is constructed. RR is determined by the attacker's choice of reference set, not sampled at random from the pool of client updates. Under the default threat model, RR is, by construction, the set of honest updates in the current round, which the omniscient adversary observes (as in ALIE, Min-Max, and Min-Sum). In the attacker-local variant, RR consists of the compromised clients' own pre-attack updates, and the adversary knows which clients it controls. In both cases the attacker does not place malicious updates in RR, so no honest/malicious filtering step is required, and GRASP does not perform honest/malicious classification on the submitted updates.

Handling honest non-IID heterogeneity. Honest non-IID updates are therefore part of RR by construction under the default threat model. Their dispersion sets the radius τ\tau (a percentile of the pairwise distances within RR), so a more heterogeneous honest population yields a wider radius: as the Dirichlet concentration decreases from α=0.5\alpha=0.5 to 0.050.05, the normalized radius τ/∥u∥\tau/\|u\| grows from 1.05 to 1.09 under Multi-Krum and from 1.27 to 1.36 under FLAME, while the honest distance distribution stays light-tailed (D90/D50=1.30D_{90}/D_{50}=1.30–1.451.45; heterogeneity table). GRASP then selects the largest scale that keeps the crafted update within τ\tau of at least qq references, so that together with the mm malicious submissions it reaches a geometric majority (Proposition 3).


Q2. Can the method be reversed into a defense, or used to improve defenses?

This is a valuable direction, and GRASP already offers a defender both a stress test and two concrete implications for defense design.

What GRASP already provides. As an attack-agnostic wrapper, GRASP is directly usable as a stress test for robust aggregation: it exposes where a given rule accepts localized, high-influence modifications that whole-update geometry does not flag.

A defense hypothesis it suggests. Our results point to two directions. First, geometric support does not imply task safety: GRASP stays within the majority-support region yet still causes damage, so proximity to many client updates is not sufficient evidence that an update is benign. Second, the attack concentrates in a small set of influential blocks, which suggests inspecting those blocks specifically rather than scoring the whole update, and combining block importance with a separate measure of in-block anomaly.

Why this is not a drop-in reversal. A high Shapley score marks a block as influential, not malicious, and honest non-IID updates may also concentrate in the same influential blocks. A defense built along these lines would therefore need validation for false positives on heterogeneous honest clients and for robustness to adaptive attackers; implementing and validating such a defense is left to future work.

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

GRASP is a wrapper that strengthens existing untargeted model-poisoning attacks (Sign Flipping, ALIE, IPM, Min-Max, Min-Sum) in federated learning. It concentrates the attack in a few parameter blocks to evade defenses that judge whole updates by distance or cosine similarity

Critical strengths
  • Clean formulation: The attacks relies on two explicit decisions: which blocks to poison and how far to scale. This improves on Pill’s technique for augmenting standard Byzantine attacks.
  • Scale selection with Theorem 4, Proposition 3 under the omniscient model.
  • Extensive ablations: Random and magnitude block selection, no scaling and Pill-style, scoring location (target vs. global vs. anchor)...
  • Reproducibility details and reporting of runtime limits and hyper-parameter sensitivity.
Critical weaknesses
  1. No coordinate-wise defenses: Coordinate-wise trimmed mean (CWTM) and coordinate-wise median (CWMed) are not evaluated. Yet CWTM with NNM is provably order-optimal under heterogeneity (Allouah et al., 2023), which makes it the most relevant baseline for this setting.

  2. Missing pre-aggregation: Under non-IID data, robust aggregators are used with pre-aggregation such as Nearest Neighbor Mixing (NNM) or bucketing with standard combinations like NNM with CWTM, NNM with CWMed, NNM with RFA (geometric median) and bucketing.

  3. Inconsistent defense coverage: defenses such as Centered Clipping appear only in the MNIST experiments. They are dropped from the Fashion-MNIST, CIFAR-10 and other experiments.

  4. Limited scope/contribution:

  • GRASP is built to evade vector-level (distance or cosine-based) defenses. So, choosing GRASP only makes sense if the attacker knows the server is not using a coordinate-wise rule. This is a partial white-box assumption about the aggregator, and the paper never states it.

  • GRASP concentrates the residual in a few blocks and extrapolates. This produces per-coordinate outliers, which are exactly what coordinate-wise defenses remove. Against CWTM or CWMed, GRASP may perform worse than the base attack it wraps.

  • The paper scopes the blind spot to defenses that assess complete updates and concludes that defenses should inspect at finer granularity. Yet coordinate-wise rules (CWTM, CWMed), which already operate at that level are never evaluated.

  • The contribution is incremental and limited: GRASP follows the same structure as Pill replacing Pill’s heuristics with Shapley-based selection and majority-aware scaling. As a wrapper, its effect depends entirely on existing standard attacks, and its gains over strong base attacks are limited despite substantial added complexity.

  • The default adversary is omniscient, in contrast, PoisonedFL (Xie et al., CVPR 2025) is a non-omniscient attack that requires no knowledge of benign clients’ updates or of the server’s aggregation rule, yet remains effective against state-of-the-art defenses (not included in the paper).

Questions for authors
  • How does GRASP perform against CWTM, CWMed and RFA, with and without NNM or bucketing ?

  • Why is CIFAR10 tested under IID data ?

4 comments
AuthorsOpenReview ↗

Response to Reviewer NHGR (1/4): Summary of New Results; W1 – Coordinate-wise Defenses (CWTM/CWMed)

We thank the reviewer for the detailed and constructive comments. We address each point below. Revisions are marked in blue; new experimental results are provided in the anonymous repository and will be incorporated into the revision (new appendix). Throughout, Attack Impact (AI) denotes clean accuracy minus attacked accuracy, in percentage points (pp).

Summary of new results

  • We add the requested baselines: CWTM, CWMed, RFA (standalone and with NNM), and PoisonedFL as a sixth base attack. All eleven defenses are now evaluated on all three datasets (198 FedSGD settings). GRASP exceeds the base attack in 180/198 settings and Pill in 176/198 (mean AI: Base/Pill/GRASP = 15.86/16.89/27.84; W4d).
  • CWTM/CWMed lie outside the whole-update scope of the paper, now defined explicitly. Standalone, GRASP stays close to Base (−0.67 / +1.17 pp) and exceeds Pill in 9/10 settings; with NNM pre-aggregation, a whole-update step, it strengthens 14/15 settings.
  • The attacker-knowledge assumption is now stated in the Threat Model; with the attacker-local variant, PoisonedFL+GRASP requires no honest-client updates.
  • CIFAR-10 is additionally evaluated under non-IID data; GRASP strengthens all 20 tested pairs (Q2).

W1. No coordinate-wise defenses (CWTM, CWMed)

Scope clarification. By whole-update defenses we mean rules that accept, reject, reweight, or clip each client's update as a unit according to a summary score over the complete update: a distance (Multi-Krum, FLDetector), a similarity (FLTrust, FLAME), or sign statistics / coordinate-subset projections (SignGuard, DnC). CWTM and CWMed aggregate each coordinate independently and make no client-level decision, so they are outside this class. In the revision we state this definition in the Introduction and discuss coordinate-wise rules and hybrids (Bulyan; NNM followed by a coordinate-wise rule) in Related Work. Whole-update criteria remain widely used, including in recent privacy-preserving robust aggregation that screens updates by cosine similarity or norm (ESFL, Zeng et al. 2025; ClusterGuard, Zhao et al. 2026).

To test this boundary, we evaluated CWTM and CWMed on MNIST (LeNet, non-IID) with all five base attacks and their Pill- and GRASP-enhanced versions:

Base attackCWTM Base+Pill+GRASPCWMed Base+Pill+GRASP
Sign Flipping9.25.06.712.716.317.4
ALIE12.25.47.716.226.222.4
IPM11.65.310.312.86.815.0
Min-Sum16.85.216.913.56.411.8
Min-Max12.210.116.923.66.117.9
Average12.396.2011.7215.7512.3716.92

Under CWTM, GRASP's mean AI is slightly below Base (12.39 → 11.72) but exceeds Pill in all five cases. Under CWMed it rises from 15.75 to 16.92, exceeding Pill in four of five. With NNM pre-aggregation, GRASP strengthens 9/10 CWTM/CWMed settings (W2). The key question is which mechanism distinguishes coordinate-wise rules from the whole-update defenses studied in the paper. We answer it with the diagnostics below.

Why coordinate-wise rules attenuate GRASP: a weight-based diagnostic. A natural hypothesis is that CWTM/CWMed simply discard GRASP's submissions. To test this, we measure how much weight the aggregate places on malicious coordinates. At each coordinate the aggregate is a convex combination of the n = 50 submitted values; the malicious residual weight (MRW) is the total coefficient assigned to the 10 malicious submissions, averaged over coordinates, rounds, and three seeds. Under CWTM (β = 0.2) each retained value receives 1/(n−2b); under CWMed the two middle values each receive 1/2. MRW = 20% means malicious values are rank-indistinguishable from honest ones; lower values indicate suppression. For GRASP we split MRW into selected blocks and their complement (which copies the anchor). Averages over the five attacks (%):

DefenseBasePillGRASPGRASP: selectedGRASP: complement
CWTM17.9113.5525.5415.0526.73
CWMed22.2413.6927.2013.0828.81

The diagnostic rejects the "discarded submissions" hypothesis: GRASP's submissions are less suppressed than Pill's under both rules (higher MRW and higher mean AI). Higher weight alone, however, does not imply a stronger attack: under CWTM, GRASP has higher MRW than Base but lower mean AI. The breakdown explains why: the selected blocks carrying the residual are suppressed (below the 20% neutral level), while the high overall MRW comes from the anchor-copying complement, which carries no residual. Coordinate-wise rules thus attenuate exactly the coordinates that carry the attack, consistent with our diagnosis that the failure lies at the granularity of inspection. We will add these results as a new appendix.

AuthorsOpenReview ↗

Response to Reviewer NHGR (2/4): W2 – Pre-aggregation (NNM); W3 – Defense Coverage across Datasets

W2. Missing pre-aggregation (NNM, bucketing)

Our original defense set was organized along a single dimension: each defense was evaluated as a standalone aggregation rule, following a recent systematization of Byzantine-robust FL (Zhang et al., 2025). Pre-aggregation is an orthogonal dimension, and NNM is the most relevant instance of it here, because NNM selects each client's neighborhood by full-update distance and therefore acts directly on the whole-update geometry this paper studies, particularly under non-IID data. NNM is thus the combination most likely to blunt GRASP: it applies a whole-update filter before CWTM/CWMed, which already attenuate GRASP's residual. We therefore evaluated NNM followed by CWTM, CWMed, and RFA, with Pill and GRASP applied to each base attack, and we find that the enhancement largely carries over. RFA and RFA+NNM will be added to the main evaluation on all three datasets (updated Tables 2, 4, 5 in the revision).

Summary. NNM is an effective defense against the base attacks, reducing their mean AI relative to the standalone rules by 32% (CWTM), 47% (CWMed), and 67% (RFA). GRASP retains most of its effect, with reductions of 8%, 13%, and 30% (Base → GRASP: 8.46 → 10.75, 8.32 → 14.79, 3.72 → 12.96; the smallest gain is under NNM+CWTM). It strengthens 14/15 base-attack settings across the three NNM combinations (the exception is Min-Sum under NNM+CWTM) and exceeds Pill in 14/15.

Per-attack results. The per-attack AI under RFA, NNM+CWTM, NNM+CWMed, and RFA+NNM are in the attack-effectiveness table. Under standalone RFA, GRASP strengthens 4/5 base attacks (IPM is the exception); under RFA+NNM, 5/5. Pill is stronger for Min-Sum under RFA and for IPM under RFA+NNM.

Does NNM still exclude malicious updates? Since NNM is itself a whole-update step, we investigate whether GRASP still places malicious updates inside honest mixing neighborhoods. We measure the exclusion recall: the fraction of the 40×10 potential honest–attacker neighbor connections per round that NNM excludes. Lower values mean malicious updates enter honest mixing neighborhoods more often. Values are averaged over rounds and seeds; per-attack results are in the exclusion-recall table. For Sign Flipping, IPM, Min-Sum, and Min-Max, GRASP lowers exclusion recall relative to Base under all three downstream rules (and below Pill in all 15 settings), consistent with its whole-update geometric construction. ALIE is the exception, already having zero exclusion recall.

Implications. NNM's neighbor selection is itself a whole-update step, so the mismatch GRASP exploits partly carries over; how much of the gap remains depends on the downstream rule. This does not contradict NNM's worst-case guarantees; it shows that order-optimal error rates do not imply negligible empirical attack impact in the non-IID settings considered here. We prioritized NNM over bucketing, which Allouah et al. (2023) show to be weaker under heterogeneity. In the revision we will add the NNM+CWTM and NNM+CWMed results and the exclusion-recall definition as a new appendix.


W3. Inconsistent defense coverage across datasets

The cross-dataset experiments originally used five of the nine defenses: FLAME, Multi-Krum, Bulyan, SignGuard, and DnC. These five operate on the current round's submissions alone, so the cross-dataset comparison holds the defender's capabilities fixed while varying the dataset and architecture. The remaining four require additional defender resources: a server-held root dataset (FLTrust, SkyMask) or cross-round state (Centered Clipping, FLDetector). Their behavior also depends on components that must be re-specified for each dataset and model, such as the size and distribution of the root dataset, or the clipping radius, momentum, and history window.

To provide consistent coverage, we have now evaluated the remaining four defenses, together with RFA and RFA+NNM, on Fashion-MNIST and CIFAR-10. The protocol matches the original cross-dataset experiments. The six base attacks include PoisonedFL (see W4e).

Across the 66 settings on each dataset (Tables 4 and 5), GRASP exceeds the base attack in 61 (Fashion-MNIST) and 62 (CIFAR-10), Pill in 57 and 59, and both in 55 and 57. Mean AI is 25.39 and 23.29, compared with 13.71 and 9.52 for Base and 14.55 and 11.29 for Pill. Combined with MNIST, GRASP falls below the base attack in 18 of the 198 settings; the aggregate counts are summarized in W4d.

The revision marks the added columns and rows in Tables 4 and 5. The FedAvg transfer experiment (Table 6) and the component ablations (Figure 3) retain the five-defense subset, as they test training-protocol transfer and component contributions rather than defense coverage.

AuthorsOpenReview ↗

Response to Reviewer NHGR (3/4): W4a–d – Scope and Contribution over Pill

W4. Limited scope / contribution

W4a. Implicit white-box assumption about the aggregator

The paper targets whole-update defenses, as stated in the title and in the scope definition (W1). This assumption is now stated explicitly in the Threat Model. The required attacker knowledge is class-level: GRASP does not take the exact aggregation rule or its thresholds as inputs, so it is weaker than the rule-specific knowledge assumed by aggregation-tailored attacks (Fang et al., 2020; Shejwalkar & Houmansadr, 2021), although each base attack keeps its own requirements. GRASP's usefulness does not hinge on this knowledge: applied regardless of the server's rule, it raises mean AI in five of the six additional configurations (W1, W2). Misidentifying the defense class therefore costs little, whereas correctly targeting a whole-update defense yields a mean gain of 11.98 pp across the 198 FedSGD settings.

W4b. Coordinate-wise rules may remove GRASP's per-coordinate outliers

As W1 shows, this concern holds only for standalone CWTM, where GRASP is slightly below Base; under CWMed GRASP exceeds Base, and under both rules it exceeds Pill in 9/10 settings. The MRW diagnostics show that coordinate-wise rules attenuate GRASP's concentrated residual without removing it, so attenuation alone does not determine whether GRASP is weaker than Base.

A second factor is the anchor-copying design outside the selected blocks. On those coordinates every malicious submission equals the shared honest anchor, so they carry no attack residual and contribute nothing to attack strength under any defense. Against whole-update defenses this design still pays off, because identical coordinates shrink the submission's overall distance and raise its cosine similarity to honest updates, which is precisely the geometric footprint those defenses score. Against coordinate-wise rules the payoff is much smaller: those rules rank each coordinate independently, so agreement on the non-selected coordinates only affects the rankings of coordinates that carry no residual, and it does not offset the loss of attack strength on exactly the coordinates that do. This asymmetry, rather than suppression of the malicious submissions as a whole, accounts for the gap between CWTM/CWMed and the whole-update defenses in W1.

W4c. Finer-granularity defenses were not evaluated

We have now evaluated standalone CWTM/CWMed and their NNM combinations (W1, W2). The results support, rather than contradict, the finer-granularity recommendation: coordinate-wise aggregation removes most of GRASP's advantage, while the gain partly returns when a whole-update step such as NNM precedes it.

W4d. Incremental contribution

We respectfully disagree with the characterization of GRASP's contribution as merely incremental. We build on Pill's subspace-poisoning perspective and explicitly acknowledge that connection.

Methodological contributions. (i) We formulate the granularity mismatch and characterize a condition under which a selected residual preserves attack effect more strongly than geometric magnitude (Proposition 1). (ii) We introduce block selection based on a second-order Shapley approximation evaluated at the model induced by the base attack's target update, scoring individual and within-layer cooperative block contributions and yielding an attack-conditioned subspace criterion. (iii) We formulate majority-aware scaling as a one-dimensional feasibility problem along the nearest-anchor injection ray and derive an exact interval-overlap solver; Theorem 4 establishes that it returns the largest feasible scale for the fixed reference set, radius, anchor, and residual direction.

Main-experiment evidence. Across all 198 FedSGD settings, GRASP exceeds Pill in 176/198 (88.9%), Base in 180/198 (90.9%), and both in 165/198 (83.3%). Mean AI is 27.84 for GRASP, versus 16.89 for Pill and 15.86 for Base.

Isolating the new components. The controlled ablations in Figure 3 hold the remaining construction fixed while varying one component. Averaged over five attacks and five defenses, the full method achieves AI 42.69, versus 33.51/37.95 for Random/Magnitude block selection and 34.17/32.30 for unscaled/Pill-style scaling (Pill's distance-based rule applied along the GRASP injection ray).

Gains over strong base attacks. Gains remain substantial for the strongest base attacks. For Min-Max (highest mean AI among Base), GRASP raises mean AI from 21.45 to 31.23 and exceeds Base in 31/33 settings; for Min-Sum, from 15.38 to 26.47. Taking the best base attack and the best GRASP-enhanced attack per defense, GRASP increases the maximum AI in 31/33 defense–dataset pairs (30.65 → 44.18). Where a base attack is already strong (AI ≥ 40; 20 settings), the gain is small (61.94 → 64.39) because several attacks already drive accuracy close to random guessing; there Pill drops mean AI to 46.11, whereas GRASP preserves the base attack's strength.

AuthorsOpenReview ↗

Response to Reviewer NHGR (4/4): W4e (cont.) – PoisonedFL; Q1 – CWTM/CWMed/RFA ± NNM; Q2 – Non-IID CIFAR-10

W4e. Omniscient adversary / PoisonedFL (cont. of W4)

Our base attacks were chosen as standard per-round model poisoning attacks (Baruch et al., 2019; Xie et al., 2020; Shejwalkar & Houmansadr, 2021), each constructing a target update from the current round, which is exactly the input to GRASP's block selection and scaling. PoisonedFL instead derives its effect from multi-round consistency: a fixed sign pattern across rounds with magnitude adapted from global-model feedback. Whether such an attack benefits from per-round subspace confinement is therefore not obvious, which is why it was not included originally. We have added PoisonedFL, PoisonedFL+Pill, and PoisonedFL+GRASP to all three main evaluations across all eleven defenses.

PoisonedFL results. Across the 33 PoisonedFL settings, PoisonedFL+GRASP has higher mean AI than standalone PoisonedFL in 30/33 and than PoisonedFL+Pill in 31/33. Mean AI is 24.73 for PoisonedFL+GRASP, versus 14.96 for PoisonedFL and 12.76 for PoisonedFL+Pill.

Dataset# settings GRASP > PoisonedFL# settings GRASP > PoisonedFL+PillPoisonedFL AI+Pill AI+GRASP AI
MNIST9/1111/1125.8220.0730.92
Fashion-MNIST10/1110/119.2010.4619.66
CIFAR-1011/1110/119.867.7523.61
Total30/3331/3314.9612.7624.73

The largest gains over standalone PoisonedFL are under Bulyan on MNIST (48.3 → 72.1), SignGuard on Fashion-MNIST (9.3 → 36.2), and FLDetector on CIFAR-10 (3.4 → 60.4). The exceptions are SignGuard and Centered Clipping on MNIST and Multi-Krum on Fashion-MNIST (39.0 → 30.7), where GRASP reduces AI relative to PoisonedFL.

PoisonedFL without honest-client updates. The results above use default GRASP, whose reference geometry is built from honest-client updates. To remove this assumption, we evaluate PoisonedFL on MNIST with the attacker-local variant, which replaces honest references with the compromised clients' own current-round pre-attack updates (five runs). Since PoisonedFL itself requires no honest-client updates, this pipeline uses no honest-client information. Averaged over the eleven defenses, it raises mean AI from 25.82 to 29.43, close to the 30.92 of default GRASP, and exceeds PoisonedFL in 10/11 defenses (the exception is Multi-Krum, 37.5 → 22.3). Sign Flipping and ALIE are likewise fully attacker-local under this variant: across the 33 settings of these three attacks, it raises mean AI from 22.13 to 34.85 (default GRASP: 34.89) and exceeds Base in 28/33.


Questions for Authors

Q1. How does GRASP perform against CWTM, CWMed, and RFA, with and without NNM or bucketing?

We evaluated CWTM, CWMed, and RFA both without pre-aggregation and with NNM; per-attack results are in W1 (standalone coordinate-wise rules) and W2 (RFA and all NNM combinations). The number of the five base attacks that GRASP strengthens in each configuration is:

DefenseWithout NNMWith NNM
CWTM2/54/5
CWMed3/55/5
RFA4/55/5

GRASP retains enhancement benefits with NNM, while the outcome depends on the downstream rule and the base attack. We did not evaluate bucketing, which Allouah et al. (2023) show to be weaker than NNM under heterogeneity; we will state this scope choice explicitly in the final revision.

Q2. Why is CIFAR-10 tested under IID data?

MNIST and Fashion-MNIST already evaluate GRASP under non-IID partitions (Dirichlet α = 0.1). We used IID CIFAR-10 to complement these experiments and test whether the enhancement also holds without heterogeneity, so the overall evaluation spans both distributions, three datasets, and three architectures. To isolate the effect of heterogeneity on CIFAR-10 itself, we additionally ran a paired non-IID evaluation on the same dataset and architecture. Base→GRASP AI pairs (pp, averaged over three seeds):

DefenseClean accMin-MaxSign FlippingALIEMin-SumIPM
FLAME69.7%5.1→59.73.1→59.71.9→56.53.8→59.515.0→57.7
Bulyan74.8%7.5→56.031.1→52.635.1→52.17.2→52.629.2→53.2
DnC75.0%26.7→65.015.5→64.826.7→65.027.3→65.021.4→66.6
RFA+NNM72.3%16.3→62.37.9→62.41.1→62.33.9→62.218.1→62.7

GRASP strengthens all 20 pairs. Under FLAME, DnC, and RFA+NNM it drives the model to near-random accuracy (8.4–13.2%; random guessing is 10% on CIFAR-10), and under Bulyan to 18.8–22.8%. NNM, designed for heterogeneous data, limits the base attacks (AI 1.1–18.1) but not their GRASP-enhanced versions. The enhancement therefore persists under non-IID CIFAR-10. We will add the full non-IID CIFAR-10 results as a new appendix in the revision.


We sincerely appreciate the reviewer's invaluable insights, and are grateful for the pointer to the coordinate-wise and NNM literature, which has helped us make the scope of our claim more precise.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Response to Reviewer xf5v (1/4): Summary of New Results; W1/Q1 – Attacker Access to Honest Updates

We thank the reviewer for the careful assessment and for the specific, constructive suggestions. We address all seven questions and the four weaknesses below. Revisions are marked in green; new experimental results are provided in the anonymous repository; additions described below will be incorporated into the revision. We keep our claims to untargeted poisoning and now state this scope explicitly in the revision. Throughout, Attack Impact (AI) denotes clean accuracy minus attacked accuracy, in percentage points (pp).

Summary of new results

  • Attacker-local GRASP (reference geometry from compromised clients' pre-attack updates only) nearly matches default GRASP over all 66 settings: mean AI 34.17 vs. 34.83 (Base 24.35), and exceeds Base in 52/66 (W1, Q1).
  • Heterogeneity and threshold. The radius τ, once normalized by the honest-update scale, is stable across Dirichlet α ∈ {0.05, 0.1, 0.5} (coefficient of variation 0.05–0.18); the honest distance distribution stays light-tailed D90/D50=1.30–1.45D_{90}/D_{50}=1.30–1.45. A percentile sweep shows the threshold is conservative at small p and the default p₀ = 40 is set by the majority-support proportion (W2).
  • Scale and partial participation. At N = 200 with full participation, mean AI rises from 13.11 to 32.88; under 10% and 20% client sampling it rises by 36% and 33% relative to Base, improving in 16/18 settings (W4, Q4).
  • Backdoor and representation-based defense. Against FedCC, BadNets + GRASP lowers accuracy by 43.83 pp (91.37% → 47.54%), showing the blind spot extends to a representation-based defense; GRASP does not preserve the backdoor, consistent with its untargeted, clean-loss objective (W3, Q5).
  • Collusion filters. FPD's absolute-similarity filter rejects the shared submission (MAR 0%). A per-client-anchor variant removes the duplicate signature and restores MAR to about 77–80% while keeping AI high (Q6).
  • Compute. On ResNet-18/CIFAR-10, one score-refresh round takes 14.93 s with 20 clients; memory is constant in n on GPU. Sub-sampling reference pairs for τ reduces the radius cost to O(sd) with 2.51–2.75% median relative error (Q3).

W1. Attacker access to honest updates

Q1. How does the attacker-local variant perform across the main experimental grid?

The Threat Model specifies two attacker configurations, and both are evaluated in the paper. (i) The default omniscient configuration follows the standard setting of the base attacks themselves: ALIE, IPM, Min-Max, and Min-Sum all construct their target update from honest pre-attack references (Baruch et al., 2019; Xie et al., 2020; Shejwalkar & Houmansadr, 2021). GRASP inherits exactly this knowledge and adds no new requirement, so the default experiments compare GRASP and the base attacks under identical attacker knowledge, which is the fair comparison for isolating the wrapper's effect. (ii) The attacker-local configuration (Appendix D.1) replaces honest references with compromised clients' pre-attack updates, adding no honest-update knowledge beyond what the base attack already uses; it is the lower bound on attacker knowledge and requires no assumption of omniscience.

Appendix D.2 already reports this variant on the main grid: Table 14 covers all 45 settings of five base attacks × nine defenses, with mean AI of 40.07 for the attacker-local variant, versus 40.04 for the default configuration and 27.74 for Base. The variant exceeds Base for all five attacks and exceeds the default configuration for Min-Max and ALIE; across the 45 settings it leads in 18, the default leads in 22, and Base in five. This is the direct answer to the reviewer's question: on the main grid, the attacker-local variant matches the default configuration to within 0.03 points on average, so the enhancement does not depend on honest-update access.

For the revision, we will extend this table to the 66 settings of the expanded grid (six base attacks × eleven defenses) by adding PoisonedFL and the RFA / RFA+NNM columns, so that the attacker-local results cover exactly the same settings as the updated Table 2. This broadens the coverage of the diagnostic to match the updated Table 2, without changing the 45-setting conclusion reported in Appendix D.2. The two configurations together bracket the achievable effect: the default isolates the wrapper under matched knowledge, and the attacker-local variant establishes that the enhancement survives without honest-update access.

AuthorsOpenReview ↗

Response to Reviewer xf5v (2/4): W2/Q4b/Q2 – Heterogeneity and Hyperparameter Sensitivity; W3/Q5 – Backdoor Attacks

W2. Heterogeneity and threshold sensitivity

Q4b. How sensitive are results to heterogeneity and to the percentile threshold τ?

We ran two sensitivity sweeps on MNIST (Min-Max, Min-Sum, ALIE vs. Multi-Krum and FLAME, three seeds).

(i) Data heterogeneity. We sweep the Dirichlet concentration α ∈ {0.05, 0.1, 0.5} at the default percentile p₀ = 40 (table). The raw radius τ tracks the honest-update norm during training, which is expected because τ is defined as a percentile of the reference distances (figure, panel (a)). Once normalized by the honest-update scale, τ is stable, with per-round coefficient of variation 0.05–0.07 under FLAME and 0.14–0.18 under Multi-Krum. The honest distance distribution is light-tailed across a 10× change in heterogeneity (D90/D50=1.30–1.45D_{90}/D_{50}=1.30–1.45), so a percentile threshold is not pushed out by a heavy tail. τ thus adapts to the update scale, and GRASP exceeds the base attack in both MAR and AI in all six cells.

(ii) Threshold selection. At α = 0.1 we sweep the percentile p ∈ {20, 30, 40, 50, 60} with q fixed (table). The threshold is not overly permissive: small p is conservative — at p = 20 only 14.0% (Multi-Krum) / 30.0% (FLAME) of rounds are feasible, so the method falls back to the anchor; large p injects larger residuals but is admitted less — from p₀ = 40 to p = 60, MAR falls 78.7 → 39.9 (Multi-Krum) and 89.1 → 17.3 (FLAME). The default p₀ = 40 follows the majority-support proportion in Eq. (5).

Q2. Hyperparameter sensitivity (Appendix C.7)

(i) Block budget and the impact–stealth trade-off. Table 13 varies the per-layer block selection rate over 5%, 10%, 20%, 30%, 50%, and 100% (Min-Max, Min-Sum, ALIE × Multi-Krum, SignGuard, DnC, FLAME). The default 10% gives mean AI of 31.66, close to the best tested average of 31.74 at 30%; all-block injection reduces it to 28.71. For example, Min-Sum's extrapolation rate falls from 8.2% at 10% budget to zero at budgets ≥ 30%, whereas ALIE extrapolates in 96.5% of rounds at 100% budget yet its AI falls from 24.74 to 14.90. Neither a wider mask nor more frequent extrapolation guarantees greater impact.

(ii) Shapley weights. Figure 4 sweeps αs\alpha_s with βs=0.5\beta_s=0.5, and βs\beta_s with αs=1\alpha_s=1, on MNIST/LeNet under non-IID data. Mean AI increases from 25.61 at αs=0\alpha_s=0 to 31.66 at αs=1\alpha_s=1, then falls to 28.44 at αs=2\alpha_s=2. For βs\beta_s, the best tested average is 35.93 at 0.125, versus 31.66 at the default 0.5. The attack-specific optima differ, so the defaults are shared settings rather than uniformly optimal choices.

(iii) Scoring-panel size. Table 14 varies the panel cap over 32, 64, and 128 examples (holdout fraction fixed at 10%). Across Sign Flipping, IPM, and Min-Max against FLTrust, FLAME, and Multi-Krum, mean AI is 39.78, 37.08, and 43.58, respectively. The maximum average decrease from the 128-example default is 6.50 points, although individual settings vary more (up to a 28.71-point range).


W3. Backdoor attacks and scope

Q5. Could you evaluate backdoor attacks and compare to a representation-based defense?

Scope. GRASP is a wrapper for untargeted poisoning: its block score (Eq. (3)) uses the clean-label loss, so it selects blocks that most degrade main-task utility, and nothing in it rewards preserving a trigger-to-target mapping. We will state this scope in the Threat Model (Section 2.4) and the Introduction in the revision.

Experiment. We ran BadNets (all-to-one, target 7) with and without GRASP against FedCC, a representation-based defense that compares penultimate-layer representations, and against undefended Mean (MNIST/LeNet, FedSGD, 100 rounds, α = 0.1, n = 50, m = 10, seeds 1–3; final-round ACC/ASR).

AggregationAttackFinal ACCFinal ASR
MeanBadNets94.6194.18
MeanBadNets + GRASP30.07 (−64.54)11.03
FedCCBadNets91.3760.97
FedCCBadNets + GRASP47.54 (−43.83)2.27

(i) The granularity blind spot extends to a representation-based defense. FedCC withstands BadNets alone (ACC 91.37, ASR reduced to 60.97), but with GRASP its accuracy falls by 43.83 points to 47.54. It still retains 17.47 points more than Mean, so it is not completely blind, but it is less effective against block-concentrated residuals.

(ii) GRASP is not a backdoor amplifier. GRASP lowers ASR under both FedCC (2.27) and Mean (11.03), so the drop comes from GRASP redirecting the malicious budget toward main-task degradation, as its clean-loss objective implies. A targeted extension (e.g., scoring blocks by the loss on triggered inputs) is left to future work.

AuthorsOpenReview ↗

Response to Reviewer xf5v (3/4): W4/Q4a – Scale and Partial Participation; Q3 – Runtime and Memory

W4. Scale and partial participation

Q4a. Do results hold under partial participation and larger federations?

Larger federations and partial participation. We ran MNIST/LeNet with N=200N=200 clients, four times the main setting (40 fixed malicious identities, Dirichlet α=0.1\alpha=0.1, 200 rounds, seeds 0, 1, 2). Each round the server samples C ∈ {10%, 20%, 100%} of clients uniformly without replacement, so the number of sampled attackers mtm_t varies across rounds. Defenses that need the number of attackers, as well as GRASP's support requirement qtq_t, use each round's realized ntn_t and mtm_t; rounds with mt=0m_t=0 apply no attack. Clean ACC comes from a matched no-attack run at each rate.

The full results show that GRASP improves over Base in 16 of 18 settings. With full participation at N=200N=200, mean AI rises from 13.11 to 32.88, so the enhancement carries over to a larger federation. Under partial participation the gain persists: at 10% sampling, mean AI rises from 10.06 to 13.71 (+3.65 pp, +36.3% relative to Base), and at 20% from 14.15 to 18.83 (+4.68 pp, +33.1% relative to Base); the largest single gain is DnC + Min-Max at 10% (4.04 → 12.21, roughly 3×). The absolute gains are smaller than at full participation because the model has not converged within the 200-round budget (clean ACC 67.2–75.6% at 10% and 78.8–80.5% at 20%, versus 90.5–94.9% at full participation), which reduces the accuracy headroom available to the attack. Each 10% round also samples about 4 attackers and 16 honest references, so both the coalition and the reference geometry are small and change every round. The two settings where GRASP does not improve are both Min-Max runs in which the base attack already drives accuracy low, leaving little headroom: RFA+NNM at 20% and DnC at 100%, both within 2 pp of Base.

Reference dispersion under strong non-IID (Q4). As reported in W2, the honest reference distances stay light-tailed across a 10×10\times change in heterogeneity (D90/D50=1.30D_{90}/D_{50}=1.30–1.451.45 for α∈{0.05,0.1,0.5}\alpha\in\{0.05,0.1,0.5\}), and the normalized τ\tau has a per-round coefficient of variation of 0.05–0.18, so the percentile-based τ\tau remains well-defined under strong non-IID. We did not construct a deliberately multimodal reference set and leave that to future work.


Q3. Runtime, memory, and reference sub-sampling

(i) ResNet-18/CIFAR-10 runtime. With 11.2M parameters and a 128-example scoring panel, one score-refresh round (excluding client training and aggregation) takes 14.93 s with 20 clients, as in our CIFAR-10 experiments: 5.55 s for Shapley scoring (per-example gradients and Fisher-vector products, on GPU), 1.64 s for pairwise distances/radius, 3.92 s for support counting, 2.07 s for the interval solver, and 1.75 s for the rest.

(ii) Scaling with nn and dd. Total time is 43.36 / 76.83 / 225.90 s at 50 / 100 / 200 clients. Scoring stays at ~5.5 s, while radius computation grows 99.4× (1.64 → 163.08 s) for a 10× increase in nn, consistent with its O(∣R∣2d)O(|R|^2d) cost, and becomes the main bottleneck (72.2% at n=200n=200). In dd, going from the 266K-parameter CNN to ResNet-18 (42× more parameters) raises round time only ~9.5× (1.57 → 14.93 s); Appendix C.8 (Figure 5) reports the full CNN sweeps.

(iii) Memory. Peak GPU memory is 183.32 MiB regardless of nn, since only scoring runs on the GPU. CPU RSS grows roughly linearly with nn (7.64 → 60.13 GiB from n=20n=20 to 200), dominated by stored updates and float64 temporaries in our general-purpose interval solver, which row-wise processing would reduce to O(d)O(d) working memory.

(iv) Sub-sampling references for τ\tau. Estimating τ\tau from ss uniformly sampled reference pairs reduces the radius cost from O(∣R∣2d)O(|R|^2d) to O(sd)O(sd), independent of nn. With s=100s=100, the median relative error is 2.51% at n=50n=50 and 2.75% at n=200n=200 (p95: 7.55% / 8.39%, essentially unbiased), so accuracy depends on ss rather than nn. End-to-end at n=50n=50 against Multi-Krum / FLAME, MAR drops by 3.29 / 7.00 points and feasibility by 8.00 / 4.67 points: a modest computation–performance trade-off. This is consistent with the percentile sweep in W2: across τ/τp0∈[0.91,1.08]\tau/\tau_{p_0}\in[0.91,1.08], which covers the p95 error, GRASP still clearly beats Base (AI 48.4–53.7 vs. 29.7 under Multi-Krum; 22.8–28.6 vs. 1.7 under FLAME), with over-estimation costing more admission under FLAME. A conservative (lower-confidence-bound) estimate would mitigate this.

(v) Practicality. GRASP builds one shared update per round, which an attacker-side node can compute once and distribute to the colluding clients (Sections 3 and 6). Moreover, Min-Max and Min-Sum already compute O(n2d)O(n^2d) pairwise distances among benign updates, so the wrapper does not increase their asymptotic cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.