acceptodds
Under review as a conference paper at ICLR 2027

ConCam: Anchor Consensus and Camouflaged Injection for Backdoor Attacks in Vertical Federated Learning

Abstract

Vertical federated learning (VFL) enables parties holding disjoint feature partitions to collaboratively train models without sharing raw data. However, the embedding gradients exchanged during training can be exploited by a malicious passive party to implant backdoors. Recent VFL attacks can infer target-related samples from gradient information, yet single-anchor target-set construction remains sensitive to anchor quality, while explicit triggers and relatively large poisoning budgets may introduce conspicuous attack patterns. In this paper, we propose ConCam, a targeted backdoor attack that combines anchor consensus for reliable target-set construction with camouflaged backdoor injection. Specifically, ConCam aggregates candidate target sets from multiple target-class anchors through consensus voting and employs sample-specific implicit triggers together with dynamic gradient-based sample and carrier selection to enable effective injection at low poisoning rates. Across four datasets, ConCam achieves attack success rates above 97% with at most a 1.05-percentage-point reduction in accuracy, remains effective at a poisoning rate as low as 0.06%, and achieves substantially higher attack success rates than prior attacks under dedicated VFL backdoor defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vertical Federated Learning, Backdoor Attack, Anchor Consensus, Implicit Trigger, Low-Rate Poisoning, Gradient-Based Selection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qEtSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed ConCam, a targeted backdoor attack combining anchor consensus and camouflaged backdoor injection. Multiple target-class anchors are aggregated via consensus voting to get a reliable target-set construction. Sample-specific implicit triggers and dynamic poisoning selection are done for enabling backdoor at a very low injection ratio. The proposed attack is proven effective experimentally even under some VFL defenses for backdoor attacks.

Critical strengths
  1. According to the reported results, ConCam has a strong effectiveness–accuracy trade-off. Especially in the the low-budget comparison, with only 0.06%0.06\% per-batch poisoning rate, ConCam reports 95.99%95.99\% ASR which is amazing, comparing with others.
  2. The empirical results well supported the motivation of using multiple anchors with anchor-consensus to improve performance by reducing anchor sample selection sensitivity.
  3. Multiple aspects of the attack are reported alongside ASR, including target-set purity and perceptual distortion.
Critical weaknesses
  1. I suggest adding some visualization results to the main paper, which tells the story and demonstrates the result of the method better than using cold, pale narration and numerical results.
  2. The experiments are mainly done under setting with 22 VFL parties, the active party and the malicious (passive party). What would be the case if more parties are involved in the VFL system? What if only one of the passive parties is malicious who has to rely on its own data for attack? And what if multiple passive parties collaborate to the same attack (they share the same attacking goal)? And what if multiple passive parties each has its own malicious attacking goal? I would like to see more results on these diverse settings.
  3. Another concern to me is that, if I understand correctly, the 0.060.06% poisoning rate is defined as "per epoch". However, the selected poisoning positions can change. Therefore, with the accumulation across multiple epochs, the total number of samples modified or injected with triggers can grow linearly to a significantly larger number. The actual cumulative number could be smaller, but it is not reported. Comparisons should match cumulative modifications, distinct modified records, and training duration (number of training iterations or epochs) in addition to the per-epoch rate.
Questions for authors
  1. I would suggest adding ↑,↓\uparrow, \downarrow besides the name of different evaluated statistics to give readers a direct sense of whether higher or lower values is better.
Reviewer_7cXtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ConCam, a backdoor attack against model-split based VFL, which consists of target set construction, sample-specific camoufalged injection, and dynamic key sample selection.

Critical strengths
  • The motivation is clear and the proposed method is interested.
  • The evaluation covers attack success, benign accuracy, perceptual distortion, and multiple defenses, including gradient-based defenses that substantially suppress the attack.
Critical weaknesses
  • The repeated identical BA drops in Tables 3 and 9-11 are difficult to explain as a coincidence. The BA values in Table 3 and Tables 9-11 need an explanation. For VFLIP, the drop is exactly 12.53 on CIFAR-10 (78.91 -> 66.38), CIFAR-100 (48.81 -> 36.28), and CINIC-10 (66.70 -> 54.17). The all same drop (12.53) appears on all datasets, and similar patterns occur for the other defenses. I would need to see the experiment logs, results for individual seeds, and the explanations for the same BA drops.
  • Limited practical evaluation. All experiments use two-party VFL with images partitioned into left and right halves. The practical motivation for this partitioning is insufficiently explained, and the results provide limited evidence for settings involving heterogeneous tabular or multimodal features. The authors should motivate a realistic application of spatially partitioned images, evaluate a representative non-image VFL setting, or explicitly narrow the scope of their claims.
  • All experiments are limited to two-party VFL. The authors should evaluate ConCam with more participants, such as four or eight clients, while keeping the number of attackers fixed at one. This would help establish whether the attack remains effective as the number of honest clients increases. The evaluation should also include defenses such as VFLIP and FILTER to check whether ConCam can still evade them in these settings.
  • The proposed method is quite incremental. Steganographic generation of sample-specific invisible triggers was introduced by ISSBA[1], while LFBA [2] already uses anchor-based gradient similarity, gradient-norm-based sample selection, and feature replacement. The authors should precisely identify the additional contributions of multi-anchor consensus and carrier selection, and demonstrate their benefits beyond combining established components. Particularly informative comparisons would include LFBA equipped with the same steganographic trigger generator and separate component ablations under matched anchor information and poisoning budgets. [1] Li, Yuezun, et al. "Invisible backdoor attack with sample-specific triggers." 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2021. [2] Shen, Wei, et al. "Label-free backdoor attacks in vertical federated learning." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 19. 2025.
Questions for authors
  • Why are the BA drops identical across datasets in Tables 3 and 9-11, and can you provide the experimental logs and per-seed results to explain this pattern?
  • What practical scenario motivates the simple left-right image partitioning, and does ConCam also work in a representative tabular or multimodal VFL setting?
  • With four or eight clients and only one attacker, does ConCam remain effective both without defenses and against VFLIP and FILTER?
  • What do multi-anchor consensus and carrier selection add beyond LFBA with the same steganographic trigger generator, and can separate ablations demonstrate their benefits under matched anchor information and poisoning budgets?
Reviewer_1CSQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ConCam, which achieves high-success-rate, low-poisoning-rate backdoor attacks in vertical federated learning by combining multi-anchor consensus, implicit triggers, and dynamic sample selection. The experimental results are somewhat competitive. However, given the method's reliance on known target-class anchors and per-sample gradients, as well as its notable limitations under gradient perturbation defenses, I believe the current level of novelty, experimental thoroughness, and practical applicability is insufficient to warrant acceptance.

Critical strengths

The paper addresses the problem of low-poisoning-rate, stealthy backdoor attacks in vertical federated learning. The proposed framework is clearly designed, integrating multi-anchor consensus, sample-specific implicit triggers, and dynamic key sample selection to improve the attack process from three perspectives: target sample identification, trigger stealthiness, and poisoning efficiency. The paper compares ConCam with multiple existing backdoor attack methods across four different datasets, and further analyzes the impact of poisoning rate, number of anchors, trigger stealthiness, and different defense mechanisms on attack performance. Experiments show that even when the effective poisoning rate is reduced to 0.06%, ConCam achieves an attack success rate of 95.99% on CIFAR-10 while maintaining relatively stable normal classification performance, demonstrating high poisoning efficiency.

Critical weaknesses
  • The main innovation of this paper lies in the combination and refinement of existing techniques, rather than proposing an attack mechanism with substantial novelty.
  • The Anchor Consensus scheme is essentially a voting-based combination strategy and lacks sufficient innovation or theoretical grounding.
  • The experimental section lacks necessary ablation studies and does not adequately consider heterogeneous vertical federated learning scenarios. Current experiments are mainly limited to a two-party setting with equally partitioned image features, where all parties adopt the same ResNet-18 architecture, failing to fully reflect the heterogeneity in feature distributions, feature dimensions, and model architectures encountered in real-world VFL systems.
  • The threat model and experimental settings limit the practical applicability of the proposed method, and there is a concern of overclaiming.
Questions for authors
  • Could the authors provide more fine-grained ablation studies to separately validate the individual contributions of dynamic sample selection, carrier replacement, and the implicit trigger design, and further analyze whether there are synergistic effects among these modules?
  • The paper proposes constructing the target sample set via multi-anchor voting, but has not sufficiently clarified the advantages of this mechanism over other simple gradient aggregation strategies. Could the authors add comparisons with methods such as average gradient prototyping, average cosine similarity, or weighted gradient aggregation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.