acceptodds
Under review as a conference paper at ICLR 2027

ResSub: Dual-Level Backdoor Defense via Residual Alignment and Subspace Perturbation in Vertical Federated Learning

Abstract

Vertical Federated Learning (VFL) enables collaborative learning over vertically partitioned data without sharing raw features, but remains vulnerable to backdoor attacks launched by malicious passive parties. Existing defenses primarily rely on attack detection or indiscriminate perturbation, while overlooking that backdoor behaviors fundamentally arise from the supervision and gradient feedback exposed during collaborative optimization. In this paper, we propose ResSub, a dual-level backdoor defense framework that simultaneously restructures supervision and gradient feedback in VFL. Specifically, Residual-Transformed Alignment reformulates the optimization objective of passive parties using residual-transformed targets, encouraging them to learn only the information missing from the active party's local prediction rather than complete label semantics. Meanwhile, Subspace-Structured Gradient Perturbation injects structure-aware perturbations into gradient subspaces, suppressing exploitable class-discriminative optimization signals while preserving model utility. Extensive experiments across multiple datasets and attack settings demonstrate that ResSub reduces the average attack success rates from 89.7% to 8.3%, while maintaining competitive clean-task performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
vertical federated learning, backdoor defense, trustworthy machine learning, residual alignment, gradient perturbation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wk22OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a dual-level backdoor defense framework for vertical federated learning. The method reconstructs the supervision objectives of passive parties through Residual-Transformed Alignment (RTA) and reduces class-discriminative information in gradient feedback through Subspace-Structured Gradient Perturbation (SSGP), thereby suppressing backdoor implantation. The experiments cover four datasets, two backdoor attacks, and different party settings, demonstrating good defense performance.

Critical strengths
  1. The paper investigates backdoor formation in vertical federated learning from the perspectives of supervision objectives and gradient feedback. The motivation is relatively clear, and the overall method design is reasonably justified.

  2. The experiments cover multiple datasets and party configurations. The average attack success rate decreases from 89.7% to 8.3%, while clean-task accuracy is largely maintained.

Critical weaknesses
  1. The two core components proposed in this paper are closely related to existing methods: (1) the Newton-style pseudo-residual and simplex aggregation in RTA are closely related to CKD/CLC; (2) the class-conditional Gaussian noise, symmetric KL objective, and power constraint in SSGP are extremely similar to Marvell, particularly in binary classification settings. The paper does not clarify the key differences from existing works or provide experimental comparisons.

  2. Section 4.1 claims that residual transformation can prevent passive parties from learning complete label semantics. However, according to the residual target formula, the component corresponding to the true class is always larger than those of other classes, meaning that the residual target still preserves complete class-ranking information. Further validation is needed to determine whether this transformation actually reduces label recoverability from gradient feedback.

  3. Although the paper acknowledges that adaptive attacks are not considered, I would like to know whether this is because the experiments were not completed due to time constraints, or because the proposed method fails when attackers are aware of the defense mechanisms of RTA and SSGP. This is a very important issue.

  4. In Table 2, LapNoise and SSGP differ simultaneously in noise distribution, perturbation budget, and adaptive mechanism. Therefore, the contribution of the subspace structure cannot be isolated. In addition, the baselines use fixed hyperparameters, and utility–defense curves under equal hyperparameter tuning budgets are not provided.

  5. A very strange experimental observation: in Table 1, BTI-DBF reports exactly the same BA and ASR as No Defense in 14 out of the 16 settings.

Questions for authors
  1. Could the authors specifically clarify the substantive technical differences between RTA and CLC, and between SSGP and Marvell, and experimentally validate the contributions of these differences?

  2. Given that the residual target still satisfies arg⁡max⁡jci,j=yi\arg\max_j c_{i,j}=y_i, how should the authors define "reducing complete label semantics"? Could the authors provide label inference results or other measures of information exploitability based on gradients, embeddings, or multi-round histories actually observable to passive parties, to support this mechanistic explanation?

  3. Could the authors provide results against adaptive attacks in subsequent evaluations?

  4. Could the authors include an isotropic Gaussian baseline with the same perturbation power as SSGP to isolate the contribution of the subspace structure? Meanwhile, please clarify whether all methods use consistent active-party pretraining, total training budgets, and validation-set hyperparameter tuning budgets.

  5. Could the authors explain why BTI-DBF produces exactly the same results as No Defense in so many settings in Table 1? Such a phenomenon should be impossible in a backdoor defense setting.

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

This paper proposes a backdoor defense method for vertical federated learning to defend against backdoor attacks launched by malicious passive parties. The goal is to limit the learning feedback available to attackers during backdoor injection. Specifically, the method achieves backdoor defense by restricting the supervisory signals received by passive parties (RTA), and reducing the separability of gradients from different classes (SSGP). RTA constructs residual targets based on the active party’s local predictions and the ground-truth labels, encouraging the passive party to learn the complementary information required by the active party. SSGP constructs a discriminative subspace from the class-conditional mean gradients and adds structured Gaussian noise to this subspace, thereby reducing the separability of gradients associated with different classes.

Critical strengths

1.The paper takes an interesting approach to defense in that, rather than detecting anomalies or removing backdoors, it proactively inhibits backdoor formation by having the active party reduce the leakage of supervisory signals and obscure class preferences. 2.The experiments cover diverse data types, including image, tabular, activity recognition, and multimodal data, and demonstrate excellent performance.

Critical weaknesses
  1. The experiments evaluate the defense against only two attacks, which is not sufficiently convincing. Moreover, both LFBA and LRB were designed for the original VFL feedback mechanism, whereas RTA explicitly changes the optimization targets received by the passive party. This raises questions about the fairness of the evaluation and the effectiveness of the defense against adaptive attacks. The current experimental results do not establish that a malicious passive party cannot adapt its attack once it becomes aware of the defense, and the authors also acknowledge that in limitation.
  2. The ASR results demonstrate that the method is effective against the evaluated attacks. However, the paper attributes this effectiveness to a reduction in the class-discriminative information available to the attacker without directly verifying that reduction. Relevant measurements could include label inference accuracy before and after perturbation, as well as the separability of gradients across classes. The amount of class information available to the passive party requires empirical evaluation and cannot be inferred solely from the final ASR.
Questions for authors

See weaknesses.

Reviewer_zNFrOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a defense against backdoor attacks in vertical federated learning (VFL). Passive parties can establish trigger-target associations without accessing the labels by exploiting supervision and gradient information. To address this, the authors introduce a dual defense named ResSub. It consists of two key components. First, instead of using a typical cross-entropy loss, the authors introduce a residual-based loss to guide passive parties to extract features that complement the prediction made by the active party using its own local features. Second, the authors introduce strategic perturbations to the gradients before sharing them with passive parties for local optimization. Experiments on multiple datasets demonstrate consistent improvements over existing baselines.

Critical strengths
  • The idea of reformulating the objective as residual learning is interesting, even though gradient perturbation has been studied for a long time.
  • The datasets are well chosen. Many existing studies use unrealistic datasets (e.g., artificially partitioned CIFAR-10). The VFL-specific datasets make the results more representative of how the proposed method performs in the real world.
  • The paper is well written and, overall, easy to follow.
Critical weaknesses
  • The coverage of VFL-specific literature is incomplete. Two recent studies [1, 2] have not been included. More importantly, the baseline defenses are not VFL-specific, even though VFL-specific defenses have been proposed (e.g., [1, 2, 3, 4]), and some have even been discussed in the related work section.
  • The methodology requires the active party to possess local features for prediction. It is unclear how much useful information these features need to contain.
  • Although explicitly acknowledged in the limitations section, the lack of analysis under adaptive attacks weakens the experimental evaluation. This is at odds with the third contribution, which claims extensive experiments.
  • The presentation could be improved. In particular, the distinction between supervision and gradient information is unclear at the beginning and only becomes apparent when the formulation is presented.

[1] S. Yuan, X. Li, X. Cao, H. Zhang, and R. H. Deng, “General Test-Time Backdoor Detection in Split Neural Network-Based Vertical Federated Learning,” TDSC 2025.

[2] X. Xu, Y. Zhao, Y. Han, Y. Zhu, Z. Han, G. Xu, B. Wang, S. Ji, and W. Wang, “VFLMonitor: Defending One-Party Hijacking Attacks in Vertical Federated Learning,” TIFS 2025.

[3] Y. Cho, W. Han, M. Yu, Y. Lee, H. Bae, and Y. Paek, “VFLIP: A Backdoor Defense for Vertical Federated Learning via Identification and Purification,” ESORICS 2024.

[4] P. Chen, H. Xiang, X. Du, X. Xu, X. Jiang, Z. Lu, J. Yang, Q. Duan, and W. Dou, “Universal Backdoor Defense via Label Consistency in Vertical Federated Learning,” IJCAI 2025.

Questions for authors
  1. By how much does ResSub outperform VFL-specific defenses?
  2. How important is it for the active party to possess useful features?
  3. Could the authors discuss potential adaptive attacks and how ResSub might respond?

Suggestion: Please fix the citation style.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.