acceptodds
Withdrawn from ICLR 2027

Hammer and Anvil: Toward a Theory of Backdoors in Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Learning (FL) enables distributed model training but is vulnerable to backdoor attacks, where malicious clients embed attacker-controlled behaviors into the global model. Existing defenses fail against adaptive adversaries. In this paper, we present **Hammer and Anvil** (HA), a principled theoretical framework that categorizes backdoors by the deviation, , of their updates to the mean of the updates. We identify two fundamental defense types: **Type 1 (The Anvil)**, comprising outlier detection and robust aggregation effective against large-deviation attacks, and **Type 2 (The Hammer)**, consisting of removal-based defenses effective against small-deviation attacks. We demonstrate that defenses of a single type and non-principled combined defenses inherently leave an exploitable gap for adaptive attackers. To bridge this gap, we propose the principled combination of Type 1 and Type 2 defenses. We evaluate our framework against a new, worst-case, full-information adaptive adversary that knows the benign updates, the aggregation algorithm, and its parameters, and yet this adversary fails against our combined defenses. Our empirical evaluation across various datasets and settings shows that single-typed and non-principled combined defenses are easily broken, often by a single malicious client. In contrast, our best combined defense variants, , , and , remain undefeated even in the most adversarial settings. Our results provide a principled approach for research on backdoors in federated learning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Machine Learning Security, Backdoors

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_sB5kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a theoretical framework called “Hammer and Anvil”, which is designed to defend against backdoor attacks in FL. It combines Type 1 (the Anvil: outlier detection and robust aggregation) and Type 2 (the Hammer: removal) from existing defense mechanisms. Experimental results show that even under the most adversarial settings, the proposed method remain undefeated.

Critical strengths
  1. This paper explores a critical issue: how to defend against backdoor attacks in federated learning.
  2. This paper studies the security of federated learning from both the attacker and defender perspectives.
Critical weaknesses
  1. The novelty of this paper seems limited, as it primarily combines two existing categories of backdoor defenses.
  2. The review of existing defense methods is insufficient. For example, relevant studies such as [1] and [2] are not adequately discussed. It is strongly recommended to conduct a more comprehensive review of the relevant literature.
  3. Some experimental setups appear to be unrealistic. For example, federated learning typically involves a large number of clients. It is recommended to evaluate the proposed method under more realistic experimental conditions.
  4. The theoretical analysis presented in this paper focuses primarily on δ. However, the effectiveness of defensive measures may also depend on other factors, such as attack vectors, aggregation rules, data distribution, and model architecture.

[1] “Fedrecover: Recovering from poisoning attacks in federated learning using historical information,” in 2023 IEEE Symposium on Security and Privacy.

[2] FLDetector: Defending federated learning against model poisoning attacks via detecting malicious clients. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining.

Questions for authors

To what extent does the proposed theory take into account factors other than δ? Could the related work be expanded, and could the experimental setup and implementation better reflect practical considerations?

Reviewer_s5VHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents Hammer and Anvil (HA), a framework for analyzing and mitigating backdoor attacks in federated learning. The authors classify existing defenses into two categories based on how much malicious updates deviate from benign ones: Type 1 defenses (robust aggregation and outlier detection), which address large-deviation attacks, and Type 2 defenses (backdoor removal), which address small-deviation attacks. Building on this classification, the paper argues that combining both types of defenses can close the exploitable gaps left by individual approaches. The authors further introduce a full-information adaptive adversary and propose Clipped-Super-Fine-Tuning (CSFT) as a Type 2 defense. Experiments on CIFAR-10, MNIST, and ImageNet demonstrate that several existing defenses remain vulnerable to adaptive attacks, whereas selected HA combinations substantially enhance robustness.

Critical strengths
  • The paper presents an framework for understanding the complementary strengths and limitations of different defense mechanisms. The proposed categorization provides a guiding principle for designing more robust defenses, moving beyond heuristic combinations of existing methods.

  • The paper considers a powerful full-information adaptive adversary and develops defense-specific attack strategies. The experimental results reveal vulnerabilities in several existing defenses while demonstrating substantial robustness improvements when CSFT is combined with selected Type 1 defenses, highlighting the practical potential of the proposed approach.

Critical weaknesses
  • Although the paper presents HA as a principled theoretical framework, several fundamental claims lack sufficient formalization and justification. In particular, the definition of defense robustness and the theoretical significance of the deviation parameter δ\delta remain unclear. The propositions and corollaries in Section 3 rely largely on intuitive reasoning rather than rigorous proofs, and the assumptions required for these results to hold are not explicitly stated. Moreover, the paper does not justify why the proposed two-type categorization is necessary and sufficient to characterize backdoor defenses, or whether update deviation alone is sufficient to capture attack effectiveness. These concerns weaken the paper's central theoretical contribution, especially the claim that combining the two types of defenses can eliminate the exploitable gap against adaptive adversaries.

  • While the paper surveys a wide range of existing defenses in Table 1, several proposed adaptive attacks are not directly evaluated, with their effectiveness instead inferred from experiments on other defenses. As a result, the empirical evidence does not fully support the claimed generality of the framework. More importantly, all experiments are conducted with only 20 clients, representing a relatively small-scale federated learning setting. It remains unclear whether the proposed approach would remain effective with substantially larger client populations or varying proportions of malicious participants. These limitations reduce the strength of the empirical evidence supporting the paper's broader robustness claims.

Questions for authors
  • Could the authors provide a formal definition of defense robustness and explicitly state the assumptions and threat model underlying the propositions and corollaries in Section 3? In particular, can the authors provide rigorous proofs for the claimed relationships between δ and defense effectiveness, including the existence of the thresholds δ₁ and δ₂? Additionally, how is the backdoor's expected impact Δ = mδ justified, particularly for nonlinear aggregation mechanisms?

  • Why are Type 1 and Type 2 defenses sufficient to characterize the space of backdoor defenses? Is there a formal argument establishing that these categories are exhaustive and that combining them is necessary or sufficient for robustness? Could there be defenses whose effectiveness cannot be characterized solely by update deviation δ?

  • The paper introduces the full-information adaptive adversary as a new threat model. Could the authors clarify its novelty relative to existing adaptive attacks, particularly the threat model considered by Shejwalkar and Houmansadr (2021), which the paper acknowledges as similar? What additional adversarial capabilities or insights distinguish the proposed threat model from prior work?

  • Could the authors evaluate the proposed framework with substantially more clients (e.g., 100 or 1,000) and varying malicious-client fractions? Additionally, could the authors provide experimental validation of the adaptive attacks proposed for defenses that are currently analyzed only theoretically in Table 1? These results would help establish whether the conclusions generalize beyond the selected defenses and small-scale experimental settings.

Reviewer_B2muOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Hammer and Anvil (HA), a framework for analyzing backdoor attacks and defenses in federated learning. The central idea is to characterize backdoor attacks using the expected deviation $\delta$ between malicious and benign model updates. Based on this measure, the authors categorize existing defenses into two types: Type 1 defenses, including robust aggregation and outlier detection, which are argued to be effective against large-deviation attacks, and Type 2 defenses, including backdoor removal methods, which are argued to be effective against small-deviation attacks.

The authors argue that individual defense types leave exploitable gaps and propose combining complementary defenses to reduce or potentially eliminate these gaps. They introduce Clipped Super-Fine-Tuning (CSFT) as a Type 2 defense and combine it with existing Type 1 methods. The paper also introduces a full-information adaptive adversary and develops defense-aware attack strategies.

Experiments on CIFAR-10, MNIST, ImageNet, and non-IID settings demonstrate the vulnerability of several existing defenses to adaptive attacks and report improved robustness for selected HA combinations. The authors interpret these results as empirical support for the proposed theoretical framework.

Critical strengths

S1. Important research problem and an interesting conceptual perspective.

The paper addresses an important problem in federated learning security: understanding the limitations of existing backdoor defenses against adaptive adversaries and investigating whether complementary defense mechanisms can provide stronger protection. The proposed distinction between defenses targeting large- and small-deviation attacks offers an intuitive perspective for organizing existing approaches. The attempt to move beyond individual defense mechanisms toward a more systematic understanding of their complementary properties is interesting and potentially valuable.

S2. Meaningful adaptive attack analysis and encouraging empirical results.

The paper considers a strong threat model in which the attacker has knowledge of benign updates and defense configurations, and develops tailored attack strategies against several existing defenses. The experimentally validated attacks expose meaningful vulnerabilities in commonly used aggregation and detection methods. Moreover, combining CSFT with selected Type 1 defenses substantially reduces attack success rates in the evaluated settings. These results provide useful empirical insights into the potential benefits of combining complementary defense mechanisms.

Critical weaknesses

W1. Fundamental soundness concerns regarding the theoretical framework.

My primary concern is that several central propositions and corollaries are stated with overly general applicability, without sufficiently specifying the assumptions under which they hold or providing rigorous justification. More importantly, some of these claims do not appear to hold in general.

For example, Corollary 1 claims that, for all Type 1 defenses, there exists a threshold beyond which any sufficiently large-deviation backdoor attack can be prevented. However, for norm-clipping defenses, increasing the magnitude of a malicious update beyond the clipping threshold does not change the clipped update, which may still retain a backdoor effect. Similarly, Corollary 3 claims that sufficiently small-deviation backdoors can always be removed by Type 2 defenses. This is not generally guaranteed, since benign fine-tuning may fail to modify trigger-specific parameter directions, even when the corresponding parameter deviation is small.

More fundamentally, the proposed framework primarily characterizes attacks through the magnitude of update deviation while insufficiently accounting for the direction of malicious updates, the geometry of benign update distributions, and their relationship with the model's loss landscape. These factors can substantially affect both detection and backdoor removal. Consequently, the proposed universal characterization of defense effectiveness through a single scalar δ\delta is not adequately justified.

Another critical issue concerns Proposition 3, which claims that an attack optimal against a Type 1 defense remains optimal against its combination with any Type 2 defense. This conclusion does not logically follow from Corollaries 1 and 3, even if these corollaries are assumed to hold. Maximizing attack effectiveness after aggregation is not necessarily equivalent to maximizing it after subsequent fine-tuning, since the removal stage can change the relative effectiveness of different attack directions.

**These are fundamental soundness concerns rather than minor gaps in mathematical presentation. ** They directly affect the validity of the proposed theoretical framework and the justification for evaluating the combined defense using attacks optimized against individual Type 1 defenses.

W2. Insufficient evidence supporting the central conjecture and the claimed benefits of principled defense combination.

Conjecture 1 is central to the paper's proposed framework, as it asserts the existence of Type 1 and Type 2 defenses whose effective regions overlap, thereby eliminating the exploitable gap. However, the paper provides only indirect empirical evidence for this conjecture. In particular, the experiments do not directly estimate the relevant thresholds δ1\delta_1 and δ2\delta_2 or establish that their effective defense regions overlap.

The reported robustness against a finite set of adaptive attacks is encouraging, but it does not establish that the exploitable gap has been eliminated or that the proposed combinations are robust against optimal adaptive attacks. This concern is particularly important given the unsupported optimality claim in Proposition 3.

I also have concerns about whether the experimental evaluation sufficiently establishes the distinct contribution of the proposed principled combination framework. CSFT is primarily an adaptation of an existing centralized post-training defense through gradient clipping, and its operation on an aggregated global model does not inherently exploit FL-specific information. The evaluation lacks sufficiently controlled comparisons with conventional fine-tuning and appropriately tuned SFT combined with the same Type 1 defenses. Consequently, it remains unclear to what extent the improvements are attributable to the proposed framework rather than to a conventional post-training removal procedure.

Furthermore, the empirical attack evaluation is less comprehensive than the paper's broad claims suggest. While the paper discusses adaptive attacks against 13 defenses, only seven are directly evaluated experimentally. The effectiveness of the proposed attacks against several other defenses, including MARS, MASA, BART-FL, and AlignIns, remains unverified.

Finally, the proposed defense relies on access to trusted, task-relevant benign data, an additional assumption whose practical implications are important. The reported degradation under non-IID fine-tuning conditions further limits the generality of the robustness claims.

Overall, although the empirical results are interesting, they do not provide sufficient evidence for the paper's central conjecture or its broader theoretical and security claims. Addressing these issues would require substantial theoretical and experimental revisions rather than minor clarifications during rebuttal.

Questions for authors

Q1. Theoretical assumptions and validity.

Are there additional assumptions or theoretical arguments that justify the universal claims in Corollaries 1–4 and the optimality claim in Proposition 3? In particular, what guarantees that attack effectiveness and backdoor removability can be characterized primarily by update deviation magnitude, and that an attack optimal against a Type 1 defense remains optimal after applying a Type 2 defense?

Q2. Adaptive evaluation of the complete HA pipeline.

Are the adaptive attacks used to evaluate HA explicitly optimized against the complete defense pipeline, including the subsequent CSFT stage, or are they primarily optimized against the corresponding Type 1 defenses? What existing evidence supports the claim that these attacks adequately represent the strongest adversaries against the combined defenses?

Q3. Trusted data and the distinct contribution of CSFT.

Could the authors clarify the source and construction of the trusted benign dataset, particularly whether it is drawn from the original training dataset and whether it is disjoint from the clients' local data? Also, are comparisons with conventional fine-tuning or appropriately tuned SFT under comparable settings available to distinguish the contribution of CSFT from that of standard post-training backdoor removal?

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

The paper proposes Hammer and Anvil (HA), a framework that categorizes federated backdoor defenses by their effectiveness against different malicious-update deviations, with filtering/robust aggregation for large deviations and removal for small deviations. It introduces Clipped-Super-Fine-Tuning (CSFT) and combines it with existing aggregation defenses. Evaluation against adaptive attackers with knowledge of benign updates and aggregation rules, primarily on CIFAR-10 with additional MNIST, non-IID, and limited ImageNet experiments, shows substantial improvements for selected combinations.

Critical strengths

Adding CSFT to FLAME, Krum, and Multi-Metrics reduces ASR from approximately 98–100% to 14–36% under the strongest tested CIFAR-10 setting, while retaining useful benign accuracy. These results provide meaningful evidence that aggregation defenses and subsequent model repair can complement each other.

Critical weaknesses
  • The theory does not establish the claimed robustness. It does not justify why the scalar deviation δ\delta sufficiently determines backdoor removability. Small update deviations need not imply easy removal through clean fine-tuning, particularly when parameter directions and latent representations are optimized for persistence. Proposition 3 also does not establish that an attack optimal against a Type 1 component remains optimal after CSFT. Consequently, the reported attacks do not establish security against the full adaptive strategy space.

  • The novelty is insufficiently distinguished from prior work. DeepSight (NDSS 2022) and FLAME (USENIX Security 2022) already motivate complementary filtering, influence limitation, and residual backdoor mitigation. Manipulating the Byzantine (NDSS 2021) explicitly considers attackers knowing both benign updates and the aggregation rule. These works substantially narrow HA's claimed conceptual and threat-model novelty. The paper needs to establish what additional guarantee, actionable combination criterion, or demonstrated capability its framework contributes.

Questions for authors
  • What precise assumptions justify the relationship between δ\delta and backdoor removability, and Proposition 3’s transfer of attack optimality to the complete HA pipeline?

  • Were attacks optimized against the complete CSFT process, including trigger or feature optimization? What knowledge of the trusted fine-tuning data and CSFT parameters was available to the attacker?

  • What is the specific advantage over the complementary-defense reasoning in DeepSight/FLAME and the existing full-information threat model?

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

Related papers

Discussion (0)

Sign in to comment.