acceptodds
Under review as a conference paper at ICLR 2027

CCCV: Cross-Client Consistency Verification for High-Precision Backdoor Defense in Vertical Federated Learning

Abstract

Vertical federated learning (VFL) enables parties with disjoint feature sets of the same samples to collaboratively train and infer without sharing raw data. However, VFL is vulnerable to clean-label backdoor attacks, where one or more malicious clients manipulate its representation to induce targeted misclassification while preserving benign labels and training behavior. Existing defenses can detect such attacks but often misclassify atypical benign representations, leading to high false-positive rates and utility degradation, while some methods additionally require auxiliary trusted data. Our key insight is to directly estimate the conditional probability of each client representation given the others. Backdoor representations create a mismatch with the cross-client context and thus receive low conditional likelihood. Compared with reconstruction-based methods that approximate conditional consistency through a point estimate and other defenses that rely on auxiliary trusted data or indirect anomaly signals, direct density estimation enables more precise detection with fewer false positives and less performance degradation without requiring additional trusted data. Based on this insight, we propose Cross-Client Consistency Verification (CCCV), an inference-time defense combining lightweight semantic screening with contrastively enhanced conditional normalizing flows. CCCV requires no trusted data and does not modify the optimization of the original VFL model, while supporting backdoor detection, attacker attribution, and prediction recovery.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vertical Federated Learning, Clean-Label Backdoor Attacks, Backdoor Defense

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_Fh6mOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CCCV, an inference-time defense against clean-label backdoor attacks in vertical federated learning. The key observation is a training–inference context shift: during training the trigger always appears alongside target-class representations from the other clients, but at inference it is attached to non-target samples, so the triggered sub-embedding no longer fits its cross-client context. To catch this, the server trains a lightweight local label estimator for each client and a conditional normalizing flow that models p(z^(k) | z^(−k)), strengthened with a cross-class contrastive loss. At inference, samples whose local predictions disagree with the global prediction are passed to the flows. These samples are flagged if their summed conditional NLL margins over class-specific reference thresholds exceed a set level. Detected samples are then used to identify the malicious clients via a recurring-pattern memory, and the prediction is recovered by masking those clients’ embeddings. The method needs no trusted clean data and doesn’t change the original VFL training. Experiments on five image and tabular datasets against three attacks show much higher detection precision, while keeping ASR low and accuracy close to the undefended model.

Critical strengths
  1. It's insightful to categorize backdoor attacks in VFL into three types.
  2. Sufficient ablation study.
  3. Detection precision and F1 improve substantially over prior defenses across five datasets and three attacks, while ASR stays low and clean accuracy is preserved. No trusted clean data is needed.
Critical weaknesses

Strong Weakness:

  1. The paper's motivation for requiring auxiliary trusted data is commonly shown in previous works, such as SPMC [1]
  2. There is no support for the author's claim that clean-label backdoor poisoning induces a training-inference shift in cross-client context. It would be better to provide a mathematical proof or experiment.
  3. What's the difference between SPMC and the conditional distribution that is the paper's claim? SPMC also uses the margin contribution to detect malicious clients, and the author also uses he conditional distribution to capture unusual actions.
  4. The attack setting is outdated. I suggest including more fresh attack strategies.
  5. Table 1 shows a comparison of CCCV with other defenses. However, I notice that in CIFAR-10 and CINIC-10, all the values on ASR are not the best. Could the author explain this phenomenon?

Minor Weakness:

  1. It would be better to show some VFL cases in real scenarios in the introduction section.
  2. Notes in Line 449-450 and Line 465-466 are format error (either added in the table caption or use a smaller font size).

[1] SPMC: Self-Purifying Federated Backdoor Defense via Margin Contribution, ICML'25.

Questions for authors

See the weakness above.

5 comments
AuthorsOpenReview ↗

Feedback for Reviewer Fh6m

Thank you for your valuable comments. We will respond to each comment and question in the following.

Response to W1:

We thank the reviewer for highlighting SPMC, an important recent backdoor defense method for horizontal federated learning (HFL). We have now included and discussed SPMC in the Introduction section. However, SPMC is designed for HFL, whereas our work focuses on vertical federated learning (VFL).

Specifically, in HFL, each client maintains a complete model architecture and communicates model updates to the server. In contrast, in VFL, each client maintains only a partial model and produces intermediate embeddings from its local features. These embeddings are then transmitted to the server, where they are concatenated and processed by a classifier to generate predictions. This fundamental architectural difference leads to distinct threat models and defense mechanisms.

We agree that the reliance on auxiliary trusted data has been extensively investigated in previous HFL studies. However, this requirement remains a challenge in VFL settings. For example, UBD requires clean validation data for latent-trigger synthesis and model purification. Therefore, reducing the dependence on trusted auxiliary data remains an important consideration in the design of VFL backdoor defenses.

Response to W2:

Figure B1 and Appendix B.1 provide direct empirical evidence for this training–inference shift. Figure B1 compares the conditional inconsistency score Γ=∑k−log⁡p(z(k)∣z−(k))\Gamma=\sum_k-\log p\left(\mathbf z^{(k)}\mid\mathbf z^{-(k)}\right) for training-time reference examples (gray), benign inference examples (blue), and backdoor inference examples (red). Across five datasets and three attacks, the backdoor inference distribution shifts markedly away from the training-time reference distribution. During training, triggered representations are observed with target-class contexts, whereas at inference, the same trigger is combined with non-target representations from benign clients. The resulting increase in conditional inconsistency directly supports the claimed cross-client context shift.

Response to W3:

Basically, SPMC and CCCV differ in three fundamental respects.

  1. Setting. SPMC addresses backdoor robustness in HFL, where clients train local models on different samples but share the same feature space. CCCV addresses clean-label backdoor inputs in split neural network-based VFL, where clients hold different features of the same sample.
  2. Core idea. SPMC measures how each client update contributes to the aggregated HFL model through marginal contribution. CCCV instead performs cross-client conditional prediction: it evaluates whether one client's current representation is probabilistically compatible with the representations supplied by the other clients.
  3. Mechanism. SPMC performs training-time gradient purification and robust server aggregation. CCCV performs inference-time local semantic screening, conditional-density verification, malicious-client attribution, and immediate prediction recovery for each input.

Response to W4:

We respectfully clarify that the evaluated attacks remain recent and representative VFL backdoor attacks. Our experiments include VILLAIN, BadVFL, and the SplitNN-specific Optimal Trigger attack, which cover representative trigger and poisoning mechanisms. These attacks remain standard evaluation choices in recent VFL defense research. For example, FILTER (2026) evaluates VILLAIN, BadVFL, and Optimal Trigger under closely related VFL settings. The same attacks are also considered by recent defenses, including CONFIRM, GBD, and Ma et al. (2026). Therefore, our attack selection follows current evaluation practice in the VFL backdoor-defense literature.

Response to W5:

CCCV focuses on high-precision backdoor defense while maintaining comparable defense effectiveness.

In contrast, more aggressive detection strategies may achieve lower ASR by flagging and modifying a larger number of inputs, including benign ones. Table 2 illustrates this trade-off. For example, under VILLAIN on CIFAR-10, VFLIP achieves a slightly lower ASR than CCCV (2% vs. 5%), but its detection precision is substantially lower (16.84% vs. 96.00%). Moreover, CCCV preserves higher clean accuracy (78.38% vs. 76.22%).

Therefore, the small difference in ASR comes at the cost of substantially more false positives and reduced model utility. CCCV prioritizes high detection precision while achieving the best F1 score across the evaluated attack-dataset combinations. Although more aggressive thresholds could further reduce ASR, they would compromise the precision and clean accuracy that motivate our method.

Response to minor weakness:

We thank the reviewer for these suggestions. We have added representative real-world VFL applications to the Introduction. We also corrected the formatting issues.

Reviewer_Fh6mOpenReview ↗

Further Question

  1. Regarding W1, as I mentioned, the need for auxiliary trusted data has been a common theme in previous research; therefore, the authors should consider including methods that do not require such data in their comparison, as simply presenting this as the paper’s contribution is insufficient.
  2. Regarding W4, the techniques you mentioned—such as FILTER (2026), CONFIRM, and GBD—are all advanced defensive measures; however, the attack methods described in the paper—VILLAIN (2023), BadVFL (2024), and Optimal Trigger Strategy (2023)—are all two years behind current research.
AuthorsOpenReview ↗

Response to Reviewer Fh6m’s Follow-up Questions

We thank the reviewer for the constructive and prompt follow-up.

Response to Q1:

We thank the reviewer for highlighting the importance of comparisons with methods that do not require auxiliary trusted data. Our existing defense baselines already include several methods that do not require auxiliary trusted data, including FILTER, VFLIP, and GBD. As shown in Tables 1 and 2, CCCV achieves substantially higher detection precision and F1 scores than these methods across five datasets and three attacks, while maintaining competitive classification accuracy and effective attack mitigation.

Importantly, we would like to clarify that the primary contribution of CCCV is high-precision backdoor detection through explicit cross-client conditional density estimation, while eliminating the need for auxiliary trusted data is an additional practical advantage of our framework.

Response to Q2:

Our original attack baselines, including VILLAIN, BadVFL, and Optimal Trigger, are aligned with those adopted by recent VFL defense studies. This selection ensures direct comparability with existing defenses under consistent experimental settings. To further strengthen our evaluation and address the reviewer’s concerns, we have additionally incorporated three recent VFL backdoor attacks (Appendix B.4, Tables B10–B11):

  • LFBA [1] improves target-class sample identification through gradient-guided label inference and selective sample switching. Our threat model already assumes full label knowledge, covering the label-access capability that LFBA seeks to obtain.
  • IASR [2] improves input-space poisoning through gradient-guided trigger optimization and magnitude regularization.
  • LAEO [2] optimizes malicious embeddings in the latent space while enforcing proximity to benign target-class representations.

Across all five datasets, the three recent attacks achieve 94%–100% ASR without defense, demonstrating their effectiveness. Nevertheless, CCCV consistently reduces their ASRs to below 6.5%, while maintaining detection precision of 82.35%–100% and recall above 95%.

Actually, these recent advances primarily focus on improving label inference and particular modules. While these techniques enhance particular components of backdoor construction, they do not explicitly address cross-client conditional consistency. In particular, the fundamental training–inference shift persists, and these improvements do not necessarily eliminate the cross-client conditional inconsistency exploited by CCCV.

Furthermore, we consider a stronger defense-aware adaptive attack (Appendix B.3) that explicitly optimizes conditional compatibility to evade CCCV. Across five attack strengths, CCCV reduces the adaptive attack's ASR to at most 6%, while maintaining precision above 95% and recall of at least 92%. Our theoretical analysis in Appendix B.2 further supports this observation by establishing, under the stated assumptions, a relationship between increasing target-class confidence and a larger lower bound on expected cross-client conditional inconsistency.

In summary, we sincerely appreciate the reviewer's valuable feedback. We hope these clarifications and additional experimental results address your concerns, and we would be happy to further clarify or address any remaining concerns or follow-up questions you may have.

[1] Label-free backdoor attacks in vertical federated learning. 2025.

[2] Silent Poisoning: A Stealthy and Unified Dual-Space Framework for VFL Backdoor Attacks. 2026.

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

This paper proposes CCCV, an inference-time defense against clean-label backdoor attacks in vertical federated learning (VFL). The method trains a local label estimator and a conditional normalizing flow for each client. It first identifies suspicious samples based on discrepancies between local and global predictions, and then uses conditional likelihood scores to detect cross-client inconsistencies.

Critical strengths
  1. The paper addresses the crucial practical challenge of detecting individual backdoored inputs without needing an additional trusted dataset or altering the original VFL training objective.
  2. The paper is well-organized and clearly written. The motivation, methodological design, and experimental results are presented in a coherent and accessible manner, making the work relatively easy to follow.
Critical weaknesses
  1. The novelty and empirical advantages of the proposed method over closely related prior work remain insufficiently established. VFLIP already exploits inconsistencies between the embedding of one client and those of other clients, while Ma et al. (2026), which is cited in the paper, explicitly investigates cross-participant consistency using latent masked autoencoders. However, the latter is not included as an experimental baseline. A controlled comparison under matched experimental settings is needed to demonstrate that the proposed methodological differences lead to meaningful performance improvements.
  2. The first-stage screening mechanism introduces a potential structural vulnerability. In the first stage, a sample is directly accepted if sufficiently many local classifiers agree with the global prediction, bypassing the conditional density verification stage entirely. Although this design reduces computational overhead, it also creates a potential blind spot: if a triggered sample induces agreement among local predictions, the more powerful second-stage detector is never activated. This is a structural concern rather than merely an implementation detail. The authors should report the proportion of successful backdoor samples that pass the first-stage screening without further verification and evaluate attacks specifically designed to manipulate this screening decision.
Questions for authors
  1. Could the authors provide a direct comparison between CCCV and Ma et al. (2026) under identical data partitions, attack settings, and comparable auxiliary model capacities? In particular, how much of the observed improvement can be attributed specifically to conditional density estimation, rather than increased model capacity or the two-stage screening procedure?
  2. How do detection, attribution, and recovery performance change as the training-time poisoning rate increases? Furthermore, could an attacker with knowledge of CCCV jointly optimize attack success and the detector's conditional likelihood scores to evade detection? Empirical results against such adaptive attacks would be valuable for assessing the robustness of the proposed defense.
  3. The ablation results show that using local classifiers alone leads to a high false-positive rate, while incorporating density-based verification substantially improves performance. This is a useful and convincing aspect of the evaluation. However, the paper does not provide a controlled comparison between conditional normalizing flows and simpler alternatives, such as conditional reconstruction errors or other consistency-based scores, under the same screening, thresholding, and recovery procedures. Consequently, it remains unclear how much of the improvement stems from modeling the full conditional distribution, as opposed to the overall pipeline design and threshold selection.
5 comments
AuthorsOpenReview ↗

Feedback for Reviewer PiwA

Thank you for your constructive comments and questions. We will respond to each comment and question in the following.

Response to W1 and Q1:

CCCV models cross-client consistency through explicit conditional density estimation rather than deterministic representation reconstruction. This distinction is motivated by the need to reliably distinguish malicious representations from atypical yet benign representations, particularly at low poisoning ratios.

Existing consistency-based defenses, including VFLIP and Ma et al. (2026), reconstruct one client’s representation from those of the other clients. Denoting the reconstruction model by fMAEf_{\mathrm{MAE}}, its objective can be expressed as

\quad \min_{\theta_{\mathrm{MAE}}}\mathbb E\left[\left\|\mathbf z^{(k)}-f_{\mathrm{MAE}}\left(\mathbf z^{-(k)}\right)\right\|_2^2\right].$$ Under the squared reconstruction loss, the optimal deterministic prediction approaches the conditional mean, $$f_{\mathrm{MAE}}^*\left(\mathbf z^{-(k)}\right) =\mathbb E\left[\mathbf z^{(k)} \mid \mathbf z^{-(k)}\right].$$ The resulting reconstruction error therefore measures deviation from a single point estimate. When $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$ is multimodal or heterogeneous, **an atypical but legitimate benign representation may lie far from the conditional mean and receive a large reconstruction error. CCCV instead uses a conditional normalizing flow to directly estimate the full conditional density $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$**, allowing multiple legitimate conditional patterns to be captured. We added Ma et al. (2026) as an additional baseline in Tables 1 and 2. CCCV, VFLIP, and Ma et al. are evaluated under matched experimental settings. On CIFAR-10, under BadVFL, VILLAIN, and Optimal Trigger, **CCCV achieves 88.29%, 96.00%, and 85.96% precision, compared with 4.16%, 16.84%, and 5.86% for VFLIP and 26.54%, 27.33%, and 22.06% for Ma et al., respectively**. Meanwhile, CCCV maintains high recall (98%, 96%, and 98%) and low ASR (2%, 5%, and 4%). These results demonstrate that conditional density estimation substantially reduces false positives while maintaining effective backdoor detection and mitigation. **To isolate the contribution of conditional density estimation from model capacity and the two-stage screening procedure, we further added a direct module-level comparison in Section 6.2**. Both CNF and MAE are trained on the same representations, use the same cross-client context $\mathbf z^{-(k)}$, and have comparable parameter counts. MAE measures conditional consistency through reconstruction errors, whereas CNF directly estimates $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$. The local-label estimator and screening gate are excluded from this comparison. We set the detection threshold to the minimum anomaly score among 100 backdoor examples to ensure 100% recall. **Under BadVFL, VILLAIN, and SplitBA, MAE incorrectly flags 52, 4, and 50 benign examples**, achieving detection precisions of 66%, 96%, and 67%, respectively. In contrast, **CNF achieves zero false positives and 100% precision across all three attacks**, while maintaining 100% recall. Under these controlled settings, the results demonstrate that conditional density estimation provides better separation between backdoor and atypical benign representations than point-based reconstruction, independently of the two-stage screening procedure. **Response to W2:** For a clean-label backdoor attack to succeed, the malicious representation must provide strong evidence for the target class, whereas the benign representations preserve the semantics of the original source class. Since the attacker cannot observe or control the benign clients’ representations or their example-wise local predictions, it cannot directly make them agree with the target class. The resulting local disagreement therefore makes successful backdoor samples difficult to pass through the first-stage screening. **Table A1 supports this mechanism: the local-label screening alone achieves 99%, 100%, and 99% recall against BadVFL, VILLAIN, and Optimal Trigger, respectively**. Thus, nearly all triggered samples are forwarded to conditional density verification. Any remaining ASR reflects unsuccessful prediction recovery after detection rather than bypassing the first-stage screening. For attacks specifically designed to manipulate the screening decision, the attacker would need at least $\tau_s K$ local classifiers to agree with the global target prediction for each inference sample. However, the attacker controls only its own representation and cannot observe or modify the benign representations or their local predictions. Therefore, under our threat model, the attacker cannot directly manipulate benign clients' example-wise local predictions, making systematic bypass of the first-stage gate challenging.
AuthorsOpenReview ↗

Continue...

Response to Q2:

We added two experiments to evaluate CCCV under different training-time poisoning rates in Appendix A and against a defense-aware adaptive attack in Appendix B.3.

First, we evaluate three target-class poisoning rates (10%, 25%, and 50%), corresponding to 1%, 2.5%, and 5% of the entire training dataset, respectively, on CIFAR-10 against BadVFL, VILLAIN, and Optimal Trigger. As shown in Table A5, increasing the poisoning rate strengthens the attacks, with the undefended ASR reaching 100% for all three attacks at 25% and 50% target-class poisoning rates. Nevertheless, CCCV consistently reduces ASR to at most 6%, maintaining 86.24–98.04% precision and 94–100% recall, with minimal impact on benign accuracy.

These results suggest that CCCV does not depend on a specific training-time poisoning rate. Under clean-label poisoning, the trigger is associated with target-class contexts during training. At inference, the same trigger is applied to non-target samples, while the benign clients continue to provide non-target representations. Therefore, the training-inference conditional mismatch remains as the poisoning rate changes, although its magnitude and the learned representations may vary.

Second, we introduce an adaptive variant of VILLAIN that explicitly optimizes the trigger to improve its conditional compatibility with benign clients' non-target representations. Specifically, we introduce an evasion loss:

= -\frac{1}{|\mathcal D_{\mathrm{opt}}|} \sum_{s\in\mathcal D_{\mathrm{opt}}} \log p^{(k_a)}\left( \mathbf z_s^{(k_a)}+\boldsymbol{\delta} \mid \mathbf z_s^{-(k_a)} \right),$$ where $\mathcal D_{\mathrm{opt}}$ denotes a held-out optimization set of non-target samples, and $p^{(k_a)}$ is the conditional density estimated by the trained CNF. Minimizing this loss encourages triggered representations to remain conditionally compatible with benign clients' representations, thereby attempting to evade CCCV. As shown in Table B9, across five attack strengths, the adaptive attacks achieve undefended ASRs of 71%–88%, while CCCV reduces them to 1%–6%, maintaining 95.10%–99.00% precision and 92%–99% recall. **These results suggest a potential trade-off between inducing target-class predictions and preserving conditional compatibility with non-target contexts, while demonstrating CCCV's effectiveness against this adaptive attack**. **Response to Q3:** **We added a controlled module-level comparison in Section 6.2** to directly isolate the contribution of conditional density estimation. The conditional normalizing flow (CNF) and masked autoencoder (MAE) are trained using the same representations and receive the same cross-client context $\mathbf z^{-(k)}$. MAE evaluates cross-client consistency through reconstruction errors, whereas CNF directly estimates the conditional density $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$. As shown in Table 3, **we use the weakest backdoor example (i.e., the one with the lowest anomaly score) as the detection threshold** for each method, ensuring 100% recall across 100 backdoor examples per attack. **Under BadVFL, VILLAIN, and SplitBA, MAE incorrectly flags 52, 4, and 50 benign examples**, yielding detection precisions of 66%, 96%, and 67%, respectively. **In contrast, CNF produces no false positives and achieves 100% precision across all three attacks while maintaining full recall**. These results indicate that conditional density estimation better accommodates atypical benign representations than point-based reconstruction, thereby reducing false positives.
AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer PiwA,

We would like to kindly follow up on our rebuttal. We would greatly appreciate any further feedback and would be happy to address any remaining concerns.

Thank you again for your time and consideration.

Reviewer_1YoqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CCCV, a post-inference backdoor defense for vertical federated learning. By combining local prediction consistency filtering, conditional normalizing flows, and contrastive learning, it detects representation inconsistencies to achieve malicious client attribution and prediction recovery. Extensive experiments across various datasets and attack settings demonstrate its advantages in reducing attack success rates and maintaining model utility.

However, the core technique is largely a combination of existing methods, and its distinct advantage over simpler conditional consistency strategies is not fully justified. Additionally, detection accuracy on some tabular datasets remains limited, while the reliance on dataset-specific thresholds and robustness against adaptive attacks require further validation. Overall, the problem is significant and the method shows potential, but it currently lacks sufficient technical novelty and rigorous verification of key mechanisms.

Critical strengths
  • The paper addresses a critical issue in VFL backdoor detection—the misclassification of normal and anomalous samples—and characterizes the distribution mismatch between training and inference caused by backdoor attacks from the perspective of cross-client conditional distribution consistency.
  • The method integrates lightweight semantic consistency filtering, contrastive-enhanced conditional density estimation, and malicious client attribution with prediction recovery, achieving a complete backdoor detection and mitigation pipeline without requiring additional trusted data or altering the original VFL optimization process.
Critical weaknesses
  • The novelty of the method and its distinction from existing work are not sufficiently justified. Prior VFL backdoor defenses have already leveraged cross-client representation consistency for malicious detection, yet the paper lacks direct comparisons with the most relevant consistency-based defenses, making it difficult to demonstrate the independent technical advantage of explicit conditional density modeling.
  • Although the paper considers varying numbers of participants and both image and tabular datasets, it does not systematically investigate the impact of imbalanced feature partitioning, heterogeneous underlying models, and varying client feature quality on the method.
  • The theoretical analysis provides limited support for the core defense mechanism. The theoretical derivation in the appendix relies on specific assumptions about conditional compatibility shifts, and the experimental descriptions provided therein are poorly presented and difficult to follow.
Questions for authors

The authors are encouraged to further clarify the substantive innovation of CCCV over existing cross-client consistency-based VFL backdoor defenses, and to validate the independent contribution of conditional density modeling through comparisons with more directly related methods and simple conditional anomaly detection strategies. Although the paper considers varying numbers of participants, it would be helpful to further analyze the impact of imbalanced feature partitioning, heterogeneous model architectures, and client feature quality differences on detection performance, so as to verify the method’s applicability in heterogeneous VFL scenarios. Regarding the theoretical analysis, the authors could further justify the practical validity of the key assumptions under real-world attacks and strengthen the robustness evaluation through experiments against adaptive attacks.

6 comments
AuthorsOpenReview ↗

Feedback for Reviewer 1Yoq

Thank you for your constructive comments and questions. We will respond to each comment and question in the following.

Response to W1:
Our method is motivated by a key challenge that existing VFL backdoor defenses have yet to adequately address: reliably distinguishing malicious representations from atypical yet benign representations, particularly at low poisoning ratios. Existing cross-client consistency-based defenses, including VFLIP and Ma et al., also have this limitation, as reconstruction errors can be large for both backdoor representations and benign representations that deviate from the reconstructed point.

Firstly, we compare the reconstruction mechanism with our method. We denote the reconstruction mechanism as fMAEf_{\mathrm{MAE}}, it learns

= f_{\mathrm{MAE}}\left(\mathbf{z}^{-(k)}\right), \quad \theta_{\mathrm{MAE}}^{*} = \arg\min_{\theta_{\mathrm{MAE}}} \mathbb{E}\left[ \left\| \mathbf{z}^{(k)}-f_{\mathrm{MAE}}\left(\mathbf{z}^{-(k)}\right) \right\|_2^2 \right].$$ Under the squared reconstruction loss, the optimal deterministic prediction approaches the conditional mean, $$f_{\mathrm{MAE}}^*\left(\mathbf z^{-(k)}\right) =\mathbb E\left[\mathbf z^{(k)} \mid \mathbf z^{-(k)}\right].$$ Consequently, the reconstruction error measures the distance from a single point estimate rather than the probability of the observed representation under the complete conditional distribution. When $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$ is multimodal or heterogeneous, **an atypical yet legitimate benign representation may be far from this conditional mean and therefore receive a large reconstruction error**. In contrast, our method uses a conditional normalizing flow to explicitly model the full conditional density $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$, allowing different legitimate conditional patterns to be captured and reducing false positives on atypical benign samples. To empirically evaluate this distinction, we added Ma et al. as an additional baseline and compared our method with two reconstruction-based defenses, VFLIP and Ma et al., in Tables 1 and 2. On CIFAR-10, under BadVFL, VILLAIN, and Optimal Trigger, **our method achieves 88.29%, 96.00%, and 85.96% detection precision**, substantially outperforming **VFLIP (4.16%, 16.84%, and 5.86%)** and **Ma et al. (26.54%, 27.33%, and 22.06%)**, respectively. Meanwhile, our method maintains high detection recall (98%, 96%, and 98%) and low ASR (2%, 5%, and 4%) across these three attacks. Similar precision improvements are observed across the other four datasets. These results demonstrate that explicit conditional density estimation better distinguishes backdoor representations from atypical benign ones, substantially reducing false positives while maintaining effective attack mitigation. We further added a direct **module-level comparison** between the CNF and masked autoencoder (MAE) in Section **6.2**. Given the same cross-client context $\mathbf z^{-(k)}$, MAE predicts $\mathbf z^{(k)}$ and evaluates consistency using reconstruction errors, whereas CNF directly estimates $p(\mathbf z^{(k)}\mid\mathbf z^{-(k)})$. We set the detection threshold to the minimum anomaly score among backdoor examples to ensure 100% recall. **For each attack, all 100 backdoor examples are correctly detected. However, MAE also incorrectly flags 52, 4, and 50 benign examples** under BadVFL, VILLAIN, and SplitBA, respectively, indicating considerable overlap between benign and backdoor anomaly scores. **In contrast, CNF achieves zero benign overlap across all three attacks**, improving detection precision from 66%, 96%, and 67% to 100% while maintaining full recall. These results demonstrate that conditional density estimation better accommodates diverse benign representations than point-based reconstruction.
AuthorsOpenReview ↗

Continue...

Response to W2:

To address concerns about the method’s applicability in heterogeneous VFL scenarios, we added new experiments covering imbalanced feature partitioning and heterogeneous bottom-model architectures in Appendix A.

First, we evaluate feature heterogeneity using both balanced and imbalanced feature partitions. We evaluate three feature partition ratios between malicious and benign clients: 50%/50%, 25%/75%, and 75%/25%, covering scenarios where the malicious client possesses fewer or more features than the benign client. As shown in Table A6, CCCV consistently reduces the ASR to at most 7% across all configurations and attacks, while maintaining detection precision above 86% and recall of at least 93%. These results demonstrate that CCCV remains effective under imbalanced feature partitioning.

Second, we evaluate model heterogeneity by assigning different bottom-model architectures to different clients. Specifically, we evaluate three combinations of bottom-model architectures: ResNet18+ResNet18 (R/R), ResNet18+CNN (R/C), and CNN+ResNet18 (C/R), including scenarios where the malicious client employs a stronger or weaker architecture than the benign client. As shown in Table A7, CCCV reduces the ASR to at most 6% across all configurations, with detection precision above 87% and recall of at least 94%. Notably, heterogeneous architectures achieve comparable or even better detection performance than the homogeneous baseline in several settings.

CCCV accommodates these heterogeneous settings because it learns a client-specific conditional distribution from the representations generated under each VFL configuration. Therefore, the defense can capture cross-client dependencies despite heterogeneous feature allocations and model architectures, enabling effective detection of backdoor-induced inconsistencies.

AuthorsOpenReview ↗

Continue...

Response to W3:

We appreciate this concern. We largely revise the theoretical analysis in Appendix B.2 and also conduct an adaptive-attack experiment to support our theoretical analysis in Appendix B.3.

First, our analysis clearly considers assumptions from the current problem setting and proposes Theorem 1. Specifically, Theorem 1 establishes a lower bound on the expected conditional inconsistency gap between attacked and benign representations:

\left[\Gamma^{(k_a)}\left(\mathbf z^{(k_a)};\mathbf z_s^{-(k_a)}\right)\right]-\mathbb E_{\mathbf z^{(k_a)}\sim p_{\mathrm{benign}}} \left[\Gamma^{(k_a)}\left(\mathbf z^{(k_a)};\mathbf z_s^{-(k_a)}\right)\right]\geq\log\frac{\tau(\alpha)(1-\epsilon)}{\epsilon}.$$ where $\tau(\alpha)$ represents the lower bound on the geometric-mean target-class confidence, and $\epsilon$ bounds the target-class posterior supported by benign clients. When $\tau(\alpha)>\epsilon/(1-\epsilon)$, the expected inconsistency of attacked representations is strictly greater than that of benign representations. Furthermore, Corollary 1.1 shows that, for $\alpha_2>\alpha_1$ with $\tau(\alpha_2)>\tau(\alpha_1)$, the theoretical lower bound increases by $$\log\frac{\tau(\alpha_2)}{\tau(\alpha_1)}>0.$$ This establishes that stronger target-class confidence yields a larger guaranteed lower bound on expected cross-client inconsistency, rather than guaranteeing monotonicity of the actual anomaly scores. **Additionally, we introduce an adaptive variant of VILLAIN that explicitly improves the conditional compatibility of triggered representations with benign clients' non-target contexts.** Specifically, we introduce an evasion loss: $$\mathcal L_{\mathrm{evasion}}(\boldsymbol{\delta}) = -\frac{1}{|\mathcal D_{\mathrm{opt}}|} \sum_{s\in\mathcal D_{\mathrm{opt}}} \log p^{(k_a)}\left( \mathbf z_s^{(k_a)}+\boldsymbol{\delta} \mid \mathbf z_s^{-(k_a)} \right),$$ where $\mathcal D_{\mathrm{opt}}$ denotes a held-out optimization set of non-target samples, and $p^{(k_a)}$ is the conditional density estimated by the trained CNF for the malicious client. Minimizing this loss encourages triggered representations to remain conditionally compatible with benign clients' non-target representations, thereby attempting to evade CCCV's detection. As shown in Table B9, the adaptive attacks achieve undefended ASRs of 71%–88%, while CCCV reduces them to 1%–6%, maintaining 95.10%–99.00% precision and 92%–99% recall. These results suggest a potential trade-off between inducing target-class predictions and preserving conditional compatibility with non-target contexts, while demonstrating CCCV's robustness against this adaptive attack.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We sincerely thank all reviewers for their thoughtful evaluations, constructive comments, and valuable suggestions.

To address the reviewers' concerns, we have made several major revisions to the manuscript, as summarized below:

  1. Comparison with Related Methods: We added Ma et al. (2026) as an additional baseline and conducted a direct module-level comparison between conditional normalizing flows (CNF) and masked autoencoders (MAE) to further demonstrate the advantages of explicit conditional density estimation.
  2. Training-Time Poisoning Rate: We conducted additional experiments with different training-time poisoning rates (10%, 25%, and 50% of target-class samples) to evaluate the robustness of CCCV under varying poisoning intensities.
  3. Heterogeneous VFL Settings: We extended our experiments to imbalanced feature partitioning and heterogeneous bottom-model architectures to evaluate CCCV under different client configurations.
  4. Theoretical Analysis: We substantially revised Appendix B.2 and introduced a new theorem and corollary to theoretically characterize the relationship between backdoor attack effectiveness and cross-client conditional inconsistency.
  5. Adaptive Attack Evaluation: We introduced a defense-aware adaptive variant of VILLAIN that explicitly optimizes conditional compatibility and evaluated CCCV against this attack under five different attack strengths.
  6. Additional Clarifications and Improvements: We clarified the first-stage screening mechanism, distinguished CCCV from SPMC, further justified our selection of attack baselines, explained the trade-off between detection precision and ASR, corrected formatting issues, and incorporated real-world VFL applications into the Introduction.

We believe these revisions have substantially strengthened the manuscript and provided more comprehensive theoretical and empirical support for our proposed method. We sincerely appreciate the reviewers' time and consideration and hope that our revisions and responses adequately address their concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.