acceptodds
Desk rejected at ICLR 2027

FedSHIELD: Adversarial Honeypot-Guided Behavioral Client Screening for Robust Federated Learning

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Learning (FL) enables collaborative training of the model between distributed clients without sharing raw data, providing significant privacy benefits. However, such training is highly vulnerable to backdoor attacks, where malicious clients insert hidden noise that compromises the global model while preserving apparent performance on clean data. Existing defenses struggle under realistic FL conditions, where benign updates exhibit high variability, and adaptive attackers can avoid statistical detection. To address these challenges, we propose FedSHIELD, a server side, honeypot guided behavioral client screening framework which evaluates client updates by measuring their reduction in loss on adversarially crafted honeypot samples. This approach enables direct detection of malicious updates without relying on layer wise statistics or prior assumptions about attack patterns. Suspicious updates are filtered out, while the remaining updates are clipped and perturbed with differential privacy noise to ensure robustness and privacy preservation. Extensive experiments on benchmark datasets such as MNIST, FashionMNIST, SVHN, and CIFAR-10 demonstrate that FedSHIELD achieves substantially lower backdoor attack success rates while maintaining high clean-data accuracy. FedSHIELD outperforms existing defenses across both IID and non-IID data distributions.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated Learning, Backdoor Attacks, Federated Aggregation, Secure Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Qm3ZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes FedSHIELD, a honeypot-guided defense framework for robust federated learning against backdoor attacks. FedSHIELD constructs adversarially perturbed honeypot samples to detect malicious client updates based on their impact on honeypot loss. Suspicious updates are filtered out, while the remaining updates are aggregated with gradient clipping and differential privacy noise. Experiments across multiple datasets and FL settings demonstrate its effectiveness in reducing backdoor attack success rates while maintaining clean accuracy.

Critical strengths
  1. The paper provides extensive ablation studies across different federated learning settings, demonstrating the contributions of individual components.

  2. The paper presents a relatively comprehensive mathematical analysis, including a convergence proof.

Critical weaknesses

Major Weaknesses

  1. Outdated baselines: The evaluated attacks and defense baselines are relatively outdated. More recent methods should be considered to better assess the effectiveness of FedSHIELD against advanced attacks and defenses.

  2. Limited dataset scale: The experiments are restricted to relatively small datasets, such as MNIST, FashionMNIST, SVHN, and CIFAR-10. Evaluations on larger and more challenging datasets, such as ImageNet, would strengthen the claims of generalizability.

  3. Dependence on auxiliary data: FedSHIELD requires clean, labeled server-side data for honeypot construction. This assumption may limit its applicability in realistic FL scenarios where such data are unavailable.

  4. Missing hyperparameter sensitivity analysis: Although the hyperparameter settings are provided, sensitivity analyses for key parameters, such as the attack objective weight λ and the detection threshold parameter β, are missing. Their impact on defense performance should be investigated.

Minor Weaknesses

  1. Experimental clarity: The authors should provide clearer descriptions of the baseline implementations, experimental configurations, and threat model to improve reproducibility.

  2. Citation formatting: Numerous citations are incorrectly formatted. For example, in Line 42, "... evading standard evaluation Wan et al. (2024)." should be revised to "... evading standard evaluation (Wan et al., 2024)."

  3. Figure quality: Figure 1 appears to be a rasterized image with small, difficult-to-read text. A properly scaled vector graphic would improve readability and presentation quality.

Questions for authors

See critical weaknesses.

1 comment
AuthorsOpenReview ↗

FedSHIELD Generalizes to Tiny ImageNet, β Is Robust, and λ Is the Attacker’s Parameter

We thank Reviewer for the detailed assessment. We provide new experimental results and clarifications for each major weakness.

Ddataset scale

We run FedSHIELD and four baselines on Tiny ImageNet with CBA (50 clients, 20% malicious):

DefenseASRAccOPS
FedAvg0.9530.5920
Krum0.3120.3940.338
FLTrust0.2560.4870.554
FedDLAD0.1980.5140.660
FedSHIELD0.1370.5320.755

FedSHIELD achieves the lowest ASR and highest OPS across all methods on Tiny ImageNet. The absolute ASR is higher than on CIFAR-10, as expected on a harder dataset, but FedSHIELD maintains a clear lead over all baselines. The behavioral gap captured by the honeypot mechanism generalizes to larger datasets and architectures. We will include these result in the revised paper.

Hyperparameter sensitivity (Major Weakness 4).

First, a clarification: λ is the attacker’s loss weight in the backdoor objective (Section 3.1), not a FedSHIELD hyperparameter. It controls attack strength and is not under the defender’s control. FedSHIELD’s key hyperparameter is β, the threshold sensitivity parameter τ = μ_s + βσ_s.

We ran a β sensitivity study on CIFAR-10 with CBA:

βASRAcc
0.30.0100.808
0.50.0090.811
0.70.0060.812
1.00.0090.812
1.50.0110.808

ASR stays within 0.6%–1.1% and clean accuracy within 0.4 points across β ∈ [0.3, 1.5]. FedSHIELD is robust to β selection over a wide range. We will include this analysis in the revision.

Auxiliary data dependency

FedSHIELD requires one clean public sample per class for honeypot construction. This data comes from public datasets (CIFAR-100, DeepFashion) — not from clients — and need not match the clients’ distribution. Appendix A demonstrates that cross-dataset honeypots maintain AUROC ≥ 0.978, confirming detection does not depend on distributional alignment. This assumption is shared by FLTrust and Sageflow, which also require server-side auxiliary data. FedSHIELD’s requirement is weaker than defenses that need labeled examples from the target attack class. We will clarify this explicitly in the revised threat model section.

**Outdated baselines **

We agree that comparisons with 2025–2026 defenses would better position FedSHIELD. We will survey recent literature and incorporate the most relevant methods into the revised experimental comparison.

Minor weaknesses.

We will add clearer descriptions of baseline implementations, experimental configurations, and the threat model to improve reproducibility. All citation formatting errors will be corrected in the revision. Figure 1 will be replaced with a properly scaled vector graphic.

Reviewer_veW8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the problem of backdoor poisoning in federated learning and proposes a server side solution. Specifically, honeypots placed near the server perform filtering of clients whose updates that do not contribute to improving the model, then do clipping, a standard defense against poisoning.

Critical strengths

While the authors have tried to present a detailed evaluation, I am not convinced by the importance of the problem and the scientific contribution over previous work.

Critical weaknesses

The problem studied by the paper has been extensively studied and the contribution of the paper is very limited.

The paper does not cite Network-Level Adversaries in Federated Learning. Giorgio Severi, Matthew Jagielski, Gokberg Yar, Yuxuan Wang, Cristina Nita-Rotaru, Alina Oprea. In IEEE CNS 2022. The paper proposes exactly the solution they describe, clients are filtered based on the contribution and clipping is performed to withstand both targeted dropping and model poisoning attacks.

Questions for authors

How is your solution different from that presented in Network-Level Adversaries in Federated Learning. Giorgio Severi, Matthew Jagielski, Gokberg Yar, Yuxuan Wang, Cristina Nita-Rotaru, Alina Oprea. In IEEE CNS 2022 ?

Why would your scheme be adopted in practice vs schemes like Krum or the one from the paper above?

2 comments
AuthorsOpenReview ↗

FedSHIELD’s Contribution and Its Distinction from Severi et al. (2022) and Krum

We thank the reviewer and will add Severi et al. (2022) to our related work. We address the novelty concern and both questions below.

Contribution. FedSHIELD contributes attack-agnostic behavioral screening: a server-side PGD-perturbed honeypot mechanism that identifies and excludes malicious clients with no knowledge of the trigger pattern, target label, or the number of compromised participants. The core insight is that adversarial perturbation places samples in loss where only a backdoor-optimized update produces disproportionate loss reduction; benign updates trained on clean objectives do not. This behavioral gap drives a principled suspicion score sk=max⁡(0,Lref−Lk)s_k = \max(0, L_{\text{ref}} - L_k), thresholded adaptively at τ=μs+βσs\tau = \mu_s + \beta\sigma_s to accommodate heterogeneous client populations without manual tuning. No labeled data from the target class is required at any point. Under adaptive attackers who know the defense, FedSHIELD holds ASR ≤ 0.1% across all four datasets while FLTrust and Sageflow reach 41–52% ASR at identical auxiliary data budgets (Table 2). It maintains AUROC ≥ 0.978 under non-IID distributions (Dirichlet α = 0.5), where most existing defenses degrade.

How FedSHIELD differs from Severi et al. The two works target different attack surfaces. Severi et al. study network-level adversaries who drop traffic to degrade availability. FedSHIELD addresses model-poisoning backdoor attacks (CBA, DBA, SPA), in which clients communicate normally but implant a hidden trigger. Severi et al. explicitly frame model poisoning as an amplification of dropping (Section II-B), not a primary mechanism. Their defense, UpSampling (Algorithm 2), reweights the sampling probability of high-contributing clients; no client is excluded. FedSHIELD excludes clients via suspicion score sk=max⁡(0,Lref−Lk)s_k = \max(0, L_\text{ref} - L_k) on PGD-perturbed honeypots. Their detection signal (Algorithm 1) requires a server-held dataset D∗D^* of the known target class: the server must already possess representative data from the class under attack. FedSHIELD needs no such knowledge; the adversarial perturbation makes the honeypot diagnostic.

Why adopt FedSHIELD over Krum or Severi et al. Krum selects one client per round by gradient proximity and requires the defender to know the number of malicious clients. Under adaptive attackers who mimic benign gradient norms, Krum’s ASR exceeds 60% (Table 2). Severi et al.’s UpSampling cannot detect backdoor attacks: a client whose updates are backdoor-optimized but gradient-norm-similar to benign clients passes its filter entirely. FedSHIELD’s honeypot scoring closes that gap: only a jointly backdoor-optimized update produces disproportionate loss reduction on the perturbed batch, while benign updates trained on clean objectives do not. The adaptive threshold τ=μs+βσs\tau = \mu_s + \beta\sigma_s adjusts to heterogeneous client populations, reducing false positives under non-IID data where fixed-threshold defenses over-exclude benign clients.

Regarding the reviewer's characterization. “Clients are filtered based on contribution and clipping is performed” conflates two distinct signals. In Severi et al., contribution measures loss improvement on a known target class. In FedSHIELD, it measures behavioral response to adversarially crafted perturbations, a signal built to expose backdoor-optimized updates. Severi et al.’s defense is UpSampling (re-weighting), not filtering (exclusion), so even the structural description does not match.

The two papers are complementary: one addresses network-level availability attacks; the other addresses model-level backdoor integrity attacks.

Thank you for your answers. I will maintain my score.

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

The author presents FedSHIELD, a server-side defense against backdoor attacks in federated learning. The proposed method generates adversarially perturbed honeypot samples using projected gradient descent (PGD) and assesses client updates based on their ability to reduce the honeypot loss. Clients exhibiting abnormally high suspicion scores are excluded, while the remaining updates undergo clipping and aggregation with differential privacy (DP) noise. The paper further provides theoretical analyses of convergence and backdoor suppression. Extensive experiments on MNIST, FashionMNIST, SVHN, and CIFAR-10 demonstrate that FedSHIELD reduces attack success rates while maintaining competitive clean accuracy relative to existing defenses, under both IID and non-IID data distributions and across multiple attack strategies.

Critical strengths
  • Rather than relying solely on the statistical properties of model updates, FedSHIELD assesses their behavioral effects on adversarially constructed samples. This offers an alternative perspective on malicious client detection, particularly in scenarios where attackers deliberately craft updates to mimic benign behavior.

  • The paper provides a comprehensive empirical evaluation across multiple backdoor attack strategies, data distributions, and baseline defenses. The results demonstrate substantial reductions in attack success rates while largely preserving clean accuracy. The additional ablation studies and evaluation against a scoring-aware adaptive attacker further strengthen the empirical evidence supporting the effectiveness of the proposed approach.

Critical weaknesses
  • The evaluation is largely restricted to relatively simple image classification benchmarks (MNIST, FashionMNIST, SVHN, and CIFAR-10) with small CNN and VGG9 architectures. Moreover, the main experiments involve only 50 clients, a relatively modest scale for practical federated learning deployments. Although the paper reports computational overhead for up to 1,000 clients, it does not evaluate the defense's effectiveness at this scale. The attack evaluation also primarily relies on a fixed source-to-target label mapping (0 to 5), leaving its robustness against more diverse attack objectives and trigger patterns insufficiently explored. In addition, comparisons with more recent backdoor defenses would provide a stronger assessment of the proposed method. Overall, these limitations raise concerns about whether the reported improvements generalize to more complex, heterogeneous, and large-scale federated learning scenarios.

  • The proposed detection mechanism relies on the assumption that malicious updates induce disproportionately large loss reductions on PGD-generated honeypot samples. However, the paper provides limited theoretical or empirical justification for why this behavior should reliably distinguish malicious updates from benign ones across diverse backdoor attacks, particularly when the attacker's target label and trigger pattern are unknown. The connection between generic adversarial perturbations and backdoor-specific behavior remains insufficiently established. Furthermore, the role of differential privacy (DP) in the defense needs further clarification. While DP serves a distinct privacy objective, its contribution to backdoor robustness is not adequately explained or isolated. Although the ablation results suggest that DP noise further reduces attack success rates, the underlying mechanism and its trade-off with clean accuracy remain unclear. More importantly, the stated privacy guarantee is conditional on the set of retained clients and does not account for the data-dependent filtering procedure, raising questions about whether it constitutes a valid end-to-end privacy guarantee.

Questions for authors
  • Does FedSHIELD require prior knowledge of the attacker's target label or trigger pattern? If not, could the authors provide further theoretical or empirical justification for why PGD-generated honeypot samples can reliably identify malicious updates targeting arbitrary labels or triggers? Additional experiments involving diverse target labels, trigger patterns, and attack objectives would help establish whether the proposed detection mechanism is truly attack-agnostic.

  • What is the primary motivation for incorporating differential privacy (DP) noise into FedSHIELD, given its focus on backdoor defense? Could the authors conduct a controlled study varying the DP noise level, including a noise-free configuration, to systematically characterize the trade-offs among privacy, robustness, and clean accuracy? In particular, the authors should clarify why introducing DP noise leads to a lower attack success rate in Table 3 and whether this improvement is consistent across different settings. Furthermore, could the privacy analysis be extended to account for the data-dependent client filtering procedure, rather than relying solely on a conditional guarantee for the retained clients?

  • Could the authors extend the evaluation to more challenging datasets, larger model architectures, and federated learning settings with substantially more clients (e.g., 100 or 1,000)? While the current runtime analysis demonstrates the computational feasibility of scaling to larger client populations, it does not establish whether the defense maintains its effectiveness at these scales. Such experiments would provide stronger evidence of the method's robustness and practical applicability in realistic federated learning environments.

  • Could the authors broaden the empirical comparison to include more recent backdoor defense methods, particularly those introduced in 2025–2026? Although the paper evaluates FedDLAD alongside several established baselines, comparisons with additional recent approaches would provide a more comprehensive assessment of FedSHIELD's relative strengths and limitations, and better position its contributions with respect to the current state of the art.

1 comment
AuthorsOpenReview ↗

FedSHIELD Scales to 1,000 Clients, Generalizes to Tiny ImageNet, and Provides a DP Robustness Analysis

We thank Reviewer for the detailed assessment and constructive questions. We provide new experimental results for each concern.

On scale and attack diversity (CW1, Q3).

We run FedSHIELD at 50, 100, 500, and 1,000 clients on CIFAR-10 with CBA (20% malicious, partial participation for 100+ clients):

ClientsASRClean Acc
500.0060.812
1000.0100.809
5000.0210.792
10000.0430.783

ASR rises from 0.6% to 4.3% as the pool scales to 1,000 clients; clean accuracy drops 2.9 points. FLTrust and Sageflow reach 41–52% ASR at 50 clients. The adaptive threshold τ=μs+βσs\tau = \mu_s + \beta\sigma_s adjusts to the population without retuning, which explains why detection degrades only gradually.

We also ran FedSHIELD on Tiny ImageNet (200 classes, 64×64) with CBA to evaluate generalizability to a larger and more challenging dataset:

DefenseASRAccOPS
FedAvg0.9530.5920
Krum0.3120.3940.338
FLTrust0.2560.4870.554
FedDLAD0.1980.5140.660
FedSHIELD0.1370.5320.755

FedSHIELD achieves the lowest ASR and highest OPS on Tiny ImageNet, confirming that the behavioral gap captured by the honeypot mechanism generalizes beyond CIFAR-10.

We also tested four source-to-target label pairs to verify label-agnosticism:

Source→TargetASRAcc
0 → 50.0060.812
1 → 70.0090.811
3 → 80.0100.809
5 → 20.0080.811

ASR is flat across all four pairs (0.6%–1.0%). Detection does not depend on a specific label mapping.

Why PGD honeypots identify arbitrary backdoors (CW2, Q1).

PGD maximizes loss on the current global model, placing samples in high-loss regions of the input space. A backdoor-optimized update reduces loss on these samples because backdoor training forces the model to associate perturbed inputs with the target class — a generalization that extends beyond the specific trigger to adversarially constructed inputs. A benign update trained on clean data produces a gradient direction largely orthogonal to the perturbation direction and does not reduce honeypot loss disproportionately. The suspicion score sk=max⁡(0,Lref−Lk)s_k = \max(0, L_{\mathrm{ref}} - L_k) captures this behavioral gap; it arises from the structure of backdoor optimization, not from any specific trigger or label. The label diversity results above validate this claim directly. We will formalize this argument in a dedicated paragraph in the revision.

Differential privacy (CW2, Q2).

We ran a three-configuration ablation on CIFAR-10 to isolate the contribution of each layer:

ConfigurationASRAcc
No DP, no clipping0.1040.763
Clipping only, no noise0.0130.760
Full FedSHIELD0.0010.762

Gradient clipping reduces ASR from 10.4% to 1.3% (8×) by bounding the influence of any single update. DP noise reduces it further to 0.1% (13×) by disrupting residual coordinated poisoning in the aggregated update. Clean accuracy stays stable across all three configurations (76.3% → 76.0% → 76.2%), confirming that DP adds robustness at virtually no accuracy cost.

Regarding the end-to-end privacy guarantee: the reviewer correctly identifies that the filtering step is data-dependent and that the conditional guarantee does not constitute a full end-to-end DP bound. We acknowledge this as a theoretical limitation, will state it explicitly in the revision, and treat a complete compositional analysis as a direction for future work.

Recent baselines (CW1, Q4).

We will survey 2025–2026 defenses and incorporate the most relevant into the revised experimental comparison.

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

Related papers

Discussion (0)

Sign in to comment.