acceptodds
Under review as a conference paper at ICLR 2027

FedHarden: Proactive Defense Against Persistent Backdoor Attacks in Federated Learning

Abstract

Federated learning (FL) enables decentralized clients to jointly train a global model by uploading locally trained updates to a central server. However, compromised clients may inject backdoors that cause trigger-specific misclassification while preserving main-task performance. Existing defenses mainly act on client updates after local training, but may leave residual backdoor influence under sustained poisoning. This influence can accumulate across rounds when benign training does not sufficiently erase it. Motivated by the low-activity parameter subspace exploited by persistent attacks, we propose FedHarden, a detection-free FL defense that proactively robustifies the global model before each communication round. FedHarden uses OOD-derived auxiliary data to introduce competing gradients and limits server-side training to selected convolutional kernels and associated batch-normalization parameters to suppress backdoor accumulation. Experiments across diverse attacks, datasets, and model architectures show that FedHarden suppresses backdoors under sustained poisoning while maintaining comparable main-task accuracy. Specifically, in a CIFAR10 experiment with 600 poisoning rounds, it reduces backdoor accuracy to near the random-guess level (10.00%), compared with 42.44% for the strongest baseline, a reduction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor attacks, Federated Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sBi8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FedHarden, a server-side defense mechanism against persistent backdoor attacks in federated learning. Prior to each training round, the server constructs two auxiliary datasets from public out-of-distribution (OOD) images: "flood data" with random labels and "shadow data" labeled by the current model. The server then leverages gradients from these datasets to select specific convolutional kernels and associated Batch Normalization (BN) parameters for additional robustification training, followed by standard client training and norm-clipped aggregation. The core motivation is to prevent attackers from exploiting parameters that receive insufficient benign training signal.

Critical strengths
  • The focus on persistent poisoning is timely and relevant. The reported performance gains are substantial, particularly against single-client attacks, and the evaluation covers a diverse range of attack types, model architectures, client data distributions, and poisoning strengths. The proposed server-side intervention is an interesting approach because it does not rely on identifying malicious clients, which is often a bottleneck in practical FL settings.

  • The ablation studies provide useful insights into the design choices. Table 5 clearly demonstrates the advantage of kernel-wise selection over parameter-wise selection under the same parameter budget. Additionally, Table 7 effectively separates the contribution of robustification from norm clipping, preventing the attribution of all improvements to the proposed training procedure alone. The inclusion of sensitivity analysis regarding OOD data sources and server-side runtime is also valuable for assessing the practicality of the defense.

Critical weaknesses
  • There is a conceptual disconnect between the theoretical motivation and the algorithmic implementation. Definition 1 identifies "low-activity parameters" based on benign training gradients, whereas Algorithm 2 selects kernels based on their response to flood data. The paper does not establish a rigorous relationship between these two quantities. Specifically, it remains unclear whether the kernels selected by Algorithm 2 are indeed underutilized during benign training. Theorem 1 does not bridge this gap, as its bound relies on assumed malicious-update constraints and Lipschitz continuity without incorporating the specific selection procedure. While the empirical improvements are significant, they could potentially be attributed to the auxiliary training acting as a general regularizer or feature disruptor rather than specifically targeting underutilized parameters. A comparison with random kernel selection, along with measurements of benign-gradient activity before and after robustification, would be necessary to validate the proposed mechanism.

  • The evaluation of adaptive attacks is insufficient for assessing robustness against a persistent adversary. The current adaptive attacks either avoid selected kernels or increase update norms, but neither directly optimizes the attack to survive subsequent rounds of FedHarden’s randomized server-side training. Given that persistent robustness is the primary objective, a defense-aware attack that accounts for the stochastic nature of the server-side robustification is crucial. Furthermore, the client participation protocol is ambiguous. Section 5 mentions 10 participating clients out of 100, while Appendix B.1 states that malicious clients participate in every poisoning round. While these statements are compatible if malicious clients are deliberately included, this should be explicitly stated. Without a clearer participation model and a stronger defense-aware attack, it is difficult to determine how much of the reported advantage would remain against an informed persistent adversary.

Questions for authors
  1. Are malicious clients guaranteed to be selected in every poisoning round, or are they sampled from a fixed population? In Table 4, do the reported malicious-client ratios refer to the total client population or only the active participants in each round?

  2. Have you measured whether the kernels selected by Algorithm 2 correspond to parameters with low benign-gradient activity? How does the proposed selection strategy compare with random kernel selection under the same parameter budget?

  3. What information about the server-side defense is assumed to be available to the attacker? Could you evaluate an attacker that explicitly optimizes for backdoor persistence in the presence of FedHarden’s subsequent robustification steps, rather than merely avoiding selected kernels or increasing update norms?

  4. How many independent runs are used for the main results, and at which specific point within each round is backdoor accuracy measured? I noticed that the Neurotoxin blended results in Tables 2 and 3 report identical backdoor accuracies but different main-task accuracies. Additionally, Tables A9 and A10 report different effective lengths for apparently similar 400-sample settings. Are these results from independent runs or different configurations?

2 comments

Considering the upcoming deadlines, we would be immensely grateful if you could provide any additional feedback or clarifications at your earliest convenience. Your guidance is crucial for the improvement of our work, and we deeply appreciate the time and effort you have invested in reviewing our responses.

We appreciate the time and care you have already devoted to reviewing our work. As the discussion deadline approaches, we would be grateful to hear whether the additional experiments and clarifications in our latest response address your remaining concerns. Thank you again for your thoughtful feedback and consideration.

Reviewer_CfQxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FedHarden tackles persistent backdoor attacks in federated learning by using public OOD data for auxiliary training and selectively updating convolutional kernels before each round, without detecting malicious clients. Experiments are solid, covering diverse attacks, triggers, architectures, and non-IID settings, and it keeps high main-task accuracy while cutting backdoor success. But the novelty is limited since it mostly combines existing techniques like OOD-assisted training, parameter selection, and gradient clipping. The theory doesn't really link to long-term backdoor suppression, and the reliance on public OOD data plus extra server cost hurts practicality.

Critical strengths
  1. The paper addresses persistent poisoning by linking low-activity parameters to backdoor persistence, offering a useful long-term defense perspective.
  2. The server-side design is practical, requiring no malicious client identification or client-side changes while balancing defense and utility via Flood and Shadow Data.
  3. Experiments are thorough, covering diverse attacks, adaptive settings, and key ablations that support effectiveness in the evaluated scenarios.
Critical weaknesses
  1. The claims regarding the core mechanism appear somewhat overclaimed. The paper attributes the defense effectiveness primarily to the reinforcement of low-activity parameter spaces, but it has not sufficiently demonstrated the relationship between the actually selected convolutional kernels and low-activity parameters. There is also a lack of direct evidence supporting this mechanistic explanation.
  2. The technical novelty of the method is relatively limited. Training Flood Data with random labels is, in expectation, equivalent to encouraging OOD samples to produce uniform predictions, which is fairly close to existing OOD regularization approaches. The paper does not adequately distinguish this conventional regularization effect from the proposed parameter reinforcement mechanism.
  3. Experimental details and ablation studies are insufficient. The paper lacks independent ablations for components such as Flood Data and Shadow Data, as well as comparisons with simple baselines like random kernel selection. Meanwhile, sensitivity analysis on key settings such as server-side training hyperparameters is limited, leaving the actual contributions of individual components and the stability of experimental results unclear. 4.There are inconsistencies in the reported experimental results, and some experimental parameters are not fully detailed.
  4. The practical applicability of the method remains limited. FedHarden is sensitive to the source of OOD data and requires additional server-side training in each round. The paper has not adequately addressed whether the method can maintain the same level of defense effectiveness when suitable public data is unavailable or when server-side computational resources are constrained.
Questions for authors

Can you provide more baseline comparison results?

1 comment

Considering the upcoming deadlines, we would be immensely grateful if you could provide any additional feedback or clarifications at your earliest convenience. Your guidance is crucial for the improvement of our work, and we deeply appreciate the time and effort you have invested in reviewing our responses.

Reviewer_wyL3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedHarden: a server-side proactive defense against persistent backdoor attacks in federated learning. Rather than identifying malicious clients or accessing private client training data, FedHarden modifies the global model before client training in each communication round. The method uses two types of auxiliary data: Flood data and Shadow data. It selects convolutional kernels based on gradient responses and updates the selected kernels and associated BatchNorm parameters while keeping the classifier fixed. The authors evaluate the method on CIFAR-10, CIFAR-100, and EMNIST, using multiple CNN architectures and several backdoor attacks. They report that FedHarden reduces backdoor accuracy in a number of settings while maintaining reasonable main-task accuracy.

Critical strengths
  1. This paper proposes an important problem and motivation: Persistent backdoor attacks are a relevant threat in federated learning. A proactive server-side defense that does not require access to private client data or explicit identification of malicious clients has practical appeal.
  2. The experiments in this paper promising empirical results across multiple settings.
  3. This paper is clearly structured and logically organized.
Critical weaknesses

The server-side compute overhead may be material for practical FL, continuously running the method may incur significant additional costs.

Questions for authors

Please clarify whether the timings include dataset generation, repeated gradient evaluations for kernel selection, robustification optimization, and aggregation. How does this overhead scale with model size and the number of flood samples?

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Our Responses (2/2)

Individual Concerns

IC1 — Random labels versus uniform targets, component ablations, and hyperparameter sensitivity (CfQx–W2/W3, Q1/Q2)

We compare random-label Flood training with uniform-target training based on Outlier Exposure. Their expected cross-entropy objectives coincide, but their finite-sample gradients differ. At uniform predictions, the uniform-target gradient is zero, while sampled random labels retain stochastic logit-gradient signals. Both constructions achieve BA near 10% for blended backdoors, while random labels reduce semantic BA from 29.70% to 14.92%.

The loss ablations show that Flood-only training suppresses both backdoor types but lowers MA to 81.23% and 82.04%. Adding Shadow consistency raises MA to 87.92% and 89.42% while keeping BA low. Shadow-only training preserves MA but leaves BA at 70.84% and 64.50%. We also provide sensitivity results for the server learning rate, training epochs, and distance-penalty coefficient.

The alternative Flood construction is discussed in Section 4, the comparisons and sensitivity results appear in Section 5.3 (Tables 7 and 8), and the gradient analysis is included in Appendix D.2.

IC2 — Defense-aware adaptive attacks (sBi8–W2/Q3 and follow-up)

We evaluate an adaptive adversary that knows FedHarden’s algorithm and configuration and can reproduce the defense locally. It uses the current broadcast model and public auxiliary data, and approximates benign updates with its own clean data. Other clients’ actual updates and future server random draws remain unknown.

Starting from a Neurotoxin update, the attacker follows Anticipate to simulate consecutive future rounds protected by FedHarden and backpropagate their losses to refine the update. EOT accounts for randomized Flood labels, noise, and minibatch sampling. For one and three future rounds, evaluation in the actual FL pipeline yields MA of 88.01% and 88.17%, and BA of 20.93% and 21.83%, respectively.

Our follow-up explains that fewer simulated optimization steps keep attack generation computationally tractable. It also discusses Table 8’s server-side sensitivity results: under standard Neurotoxin-blended attacks, penalty weights of 0.4 and 0.8 yield BA of 12.58% and 10.00%, respectively, with MA above 87% in both settings.

The adaptive-attack results and implementation details are included in Section 5.4 and Appendix D.3 (Tables A14 and A15).

IC3 — Client participation and poisoning schedules (sBi8–W2/Q1 and follow-up)

We clarified that each round includes 10 active clients from a population of 100. All configured malicious clients participate in every poisoning round, while benign clients are sampled to fill the remaining slots. Table 4’s malicious-client ratios refer to active participants: 20%, 30%, and 40% correspond to two, three, and four malicious clients among the ten participants. This threat model maximizes their opportunities to inject malicious updates throughout the evaluation.

The main comparison experiments use poisoning rounds 431–1030. Tables A9 and A10 instead evaluate continuous poisoning over an extended horizon without the 600-round cutoff. Effective length denotes the first absolute global round at which BA reaches 35%; thus, 2104 is recorded while poisoning remains active.

The participation protocol is clarified in Section 5.1, the caption of Table 4, and Appendix B.1. The extended poisoning schedule is explained in our follow-up response to sBi8.

IC4 — Auxiliary-data requirements (CfQx–W5)

FedHarden uses public images without requiring ground-truth annotations. Using 300KRANDOM for both Flood and Shadow achieves 22.90% BA under Chameleon-blended attacks. Public natural-image datasets and general-purpose image collections reduce the data-preparation burden.

We also discuss the small-data configuration in Table A9: with 200 Flood images, MA is 89.02%, and BA first reaches 35% at global round 588. Increasing the Flood budget to 400 images delays this threshold crossing to round 1874.

The auxiliary-data requirements and computation trade-offs are discussed in Section 5.3 (Tables 6 and 8), Appendix B.7 (Table A9), and Appendix D.4.

We believe that our revisions, additional experiments, and detailed clarifications have substantially addressed the reviewers’ concerns and strengthened the manuscript. We sincerely thank the Area Chair and all reviewers for their time and thoughtful consideration.

Best regards,
Authors of Submission 29651

AuthorsOpenReview ↗

Summary of Our Responses (1/2)

Dear Area Chair and Reviewers,

We sincerely thank you for your constructive and insightful feedback. Below we provide a consolidated summary of common concerns raised across multiple reviewers, followed by individual concerns specific to each reviewer. For each point, we briefly restate our response and indicate where the corresponding changes were made in the revised manuscript, highlighted in blue. Detailed, point-by-point responses are provided in the rebuttal to each reviewer.

Common Concerns

CC1 — Connection between kernel selection and low-activity parameters (CfQx–W1/Q1; sBi8–W1/Q2)

We provide additional results connecting kernel selection to benign-gradient activity, together with warm-up measurements and a random-selection comparison. Following Definition 1, parameter activity is computed as the mean absolute per-example gradient over a clean-data probe. Across four thresholds, FedHarden covers 12.51%–26.43% of the low-activity set, compared with 8.14%–8.49% for random selection. Low-activity parameter counts decrease by 20.79%–23.64% during warm-up.

Under Neurotoxin semantic poisoning on CIFAR10 using ResNet18, FedHarden achieves 89.42% MA and 14.92% BA, compared with 87.57% MA and 26.41% BA for random selection under the same parameter budget. These results provide empirical support for the selection strategy and its connection to low-activity parameters.

The random-selection comparison is included in Section 5.3 (Table 7), and the activity and warm-up analyses are included in Appendix D.1 (Tables A12 and A13).

CC2 — Server-side overhead and computational requirements (wyL3–W1/Q1/Q2; CfQx–W5)

We provide a runtime breakdown covering Flood preparation, Shadow preparation and pseudo-labeling, robustification optimization, clipping and aggregation, and internal evaluation. Kernel selection is performed once at initialization and takes 65.11 seconds; this cost is accounted for separately from per-round timings. We also report runtime scaling with Flood sample count on ResNet18 and ResNet34.

Under the default Neurotoxin-blended configuration, reducing robustification from 200 to 50 epochs per round achieves 88.81% MA and 15.20% BA with 75% fewer optimization steps. With 400 Flood samples, this reduction lowers runtime on an RTX A6000 from 45.3 to 22 seconds per round.

The reduced-budget results are included in Section 5.3 (Table 8), and the runtime analysis is included in Appendices B.5 and D.4 (Tables A7 and A16).

CC3 — Experimental reporting and reproducibility (CfQx–W4/Q2; sBi8–Q4 and follow-up)

We clarified that the main results use one independent run per configuration, with MA and BA measured after aggregation and averaged over the final 20 global rounds. We corrected the inconsistent MA entries in Table 3 to match the corresponding Neurotoxin-blended results in Table 2; BA values are unchanged. The 400-sample entries in Tables A9 and A10 come from separate ablation runs, with effective lengths of 1874 and 1910, respectively. Their ± values describe MA variability within each run’s final 20 rounds.

In our follow-up response, we further clarified that the corresponding results in Tables 2 and 5, and Tables 2 and A11, come from separate randomized runs with matched configurations within each comparison. Each run independently performs kernel selection at initialization.

We additionally conducted three randomized trials for each standard Neurotoxin trigger setting, running through global round 1030 and averaging post-aggregation BA over the final 20 rounds:

TriggerTrial 1 BA (%)Trial 2 BA (%)Trial 3 BA (%)Mean ± SD (%)
Blended10.5211.2110.0410.59 ± 0.59
Semantic13.2015.8113.7414.25 ± 1.38

SD denotes the sample standard deviation across the three trial-level averages. These additional trials consistently retain low BA for both trigger types.

The reporting corrections and clarifications appear in Section 5.1, Table 3, and Appendix B.7. The additional randomized trials are reported in our follow-up response to sBi8.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.