acceptodds
Under review as a conference paper at ICLR 2027

When Priors Veil the Threat: Backdoor Defense via Channel Activation Inspection for Pretrained Federated Learning

Abstract

Existing defenses against backdoor attacks in federated learning aim to identify malicious clients and typically rely on the discrepancies among model gradient updates to distinguish them from benign ones. In this paper, we empirically show that when the global model is initialized with pretrained backbones, the updates similarity between benign and malicious clients can be substantially amplified, thereby rendering existing defenses ineffective. To bridge this gap, we investigate client model discrepancies from the perspective of channel activations and uncover a key observation: benign models generally exhibit consistent channel activation magnitudes across samples whereas backdoored models show abnormal activation variations. Motivated by this, we propose CAS (Channel Abnormality Score), a new malicious detection metric quantifying the abnormality within the channel activation pattern of a client model via the notion of term frequency-inverse document frequency (TF-IDF). Specifically, CAS treats channel activations as words and the feature map of each class as a document. When a client contains channels that contribute abnormally to documents as measured by CAS, the server identifies the client model as backdoored and excludes it from global aggregation. Extensive experiments show that CAS significantly outperforms existing defenses under the pretrained federated learning setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated learning, Backdoor Defense, Pretraining, Channel Activation

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

This paper observes that benign models typically exhibit consistent channel activation magnitudes across samples, whereas malicious clients exhibit substantial differences in channel activation magnitudes between clean and malicious samples. Some channels respond strongly to malicious samples but weakly to clean samples, while others exhibit the opposite pattern. Based on this observation, the paper proposes CAS. During local training, clients compute and upload the mean channel activations for each class. The server constructs channel scores using TF-IDF representations and cross-class entropy, then clusters clients using the Wasserstein distance. The cluster whose centroid has the larger norm is retained for aggregation. Experiments on four datasets under non-IID data distributions demonstrate that the method provides effective defense against four different attacks compared with baseline defenses, while incurring a relatively small loss in accuracy.

Critical strengths

1.The paper presents a novel perspective on federated backdoor defense. Instead of relying on gradient or parameter discrepancies, CAS detects malicious clients by measuring abnormal class channel activation patterns using TF-IDF and activation entropy.

2.The experimental evaluation cover multiple datasets, backdoor attacks, defense, architectures, and federated learning settings.

Critical weaknesses

1.CAS requires clients to upload both model updates and class-wise mean channel activations. However, under the paper’s threat model, malicious clients that can manipulate model updates should also be able to manipulate the accompanying activation statistics. An attacker could use local clean data or a separately trained clean model to generate benign-looking CAS statistics while uploading a backdoored model inconsistent with those statistics, thereby bypassing server-side clustering. The BatchNorm mean consistency check proposed in the appendix also appears insufficient to address this issue, since both the activation statistics and the BatchNorm running means are controlled by the client and can be jointly fabricated.

2.I have concerns about the conclusion regarding channel activation inconsistency supported by Figure 2. Figure 2(d) compares two sets of clean data on a benign model, whereas Figures 2(a)–(c) compare clean and triggered data on backdoored models. Thus, both the model state and the input conditions change simultaneously. The current results cannot distinguish whether the observed activation differences arise from the backdoor embedded in the model or merely from the presence of the trigger in the input.

3.The paper does not explain how CAS handles missing classes under highly non-IID data distributions. Because CAS requires class-wise channel activation statistics, clients without samples from certain classes may naturally exhibit concentrated, low-entropy activation patterns and consequently be misclassified as malicious.

Questions for authors

see weaknesses

7 comments
AuthorsOpenReview ↗

Response to Reviewer oWNB (1/3)

We thank reviewer oWNB for the feedback and valuable suggestions. Below we provide responses to each weakness and question. Please feel free to reach out if further clarification is needed.

Response to Critical Weakness 1:

We thank the reviewer for pointing this out. We agree with the reviewer that an adaptive attacker can bypass the BN statistics authentication by simultaneously manipulating the BN layers. However, as suggested in [1] [2], BN layers often encode a substantial amount of backdoor information to enable effective triggered behavior. That is, the training data of a backdoor task can be equivalently seen as a form of out-of-distribution data relative to that of the main task. Due to change of the training data, the estimated mean μ\mu and variance σ\sigma after the backdoor task gradually deviate from those of clean task. Thus, in order to bypass the BN authentication, attackers must confine to upload the BN statistics that are not longer capable to maintain the backdoor effect. In other words, the proposed BN authentication aims to drive the adaptive attackers into a dilemma where they either get caught by the authentication or suffer from severe backdoor performance degradation. To support our claim, we consider three forms of adaptive attackers:

  • Type-1: the attacker additionally trains a benign surrogate and uploads the CAS statistics of this model.
  • Type-2: the attacker additionally trains a benign surrogate and replaces the BN layers of the backdoored model with that of the benign one.
  • Type-3: the attacker additionally trains a benign surrogate and rescales CAS statistics of the backdoored model by the ratio between the backdoored and the benign BN running means, and uploads the rescaled statistics, while summing to the backdoored model’s BN running mean. The following table demonstrates the defense performance of CAS with the proposed BN authentication under the aforementioned adaptive attack scenarios:

Table R3-1: Performance of BN authentication against adaptive attackers in CIFAR10

AttackAdaptive TypeAccASRTPRFPR
DBAno adaptive95.2410.0199.500.50
DBA195.0710.11100.00.32
DBA293.1610.790.000.00
DBA395.4010.1999.5011.00
IBAno adaptive95.4311.7399.006.38
IBA195.1614.13100.004.30
IBA293.6612.630.000.00
IBA395.748.93100.000.33

As shown, CAS with BN authentication enabled provides effective defense service under all adaptive attack scenarios, minimizing the ASR rate down to random guess level. Other FL defenses that additionally requires local clients to upload defensive statistics (e.g., Lockdown[3], FDCR[4], FEDRACE[5]) often do not offer a practice solution for the aforementioned adaptive attackers or simply assuming that the local gradients are unbreakable. In contrast, CAS naturally provides the guaranteed robustness under adaptive attackers. In the revised manuscript, we have incorporated aforementioned justification and results in Appendix F line 1066.

AuthorsOpenReview ↗

Response to Reviewer oWNB (2/3)

Response to Critical Weakness 2:

  • "I have concerns about the conclusion regarding channel activation inconsistency supported by Figure 2. Figure 2(d) compares two sets of clean data on a benign model, whereas Figures 2(a)–(c) compare clean and triggered data on backdoored models. Thus, both the model state and the input conditions change simultaneously."

We appreciate that the reviewer asks this question. We agree with the reviewer that the observation of channel activation inconsistency supported by Figure 2 comes from two different type of data consisting D^\hat{\mathcal{D}} between the backdoor and benign models. To enhance our claim, we have replaced the Figure 2 (d) to another one (link) that instead shows the activation of a benign model using poisoned D^\hat{\mathcal{D}}. In the revised manuscript, we have updated the setup in Section 2.2 line 152 to replace the figure 2 (d) in section 2.2 line 162.

  • "The current results cannot distinguish whether the observed activation differences arise from the backdoor embedded in the model or merely from the presence of the trigger in the input"

We agree with the reviewer that the results in Figure 2 can cause ambiguity in the resulted activation differences. To address this problem, we further investigate the scenario where the malicious client submit the CAS statistics computed on clean D^\hat{\mathcal{D}} (no trigger δ\delta injected) using the backdoored model. The resulted plot (link) reveals another key finding: The benign model’s channels form a compact cluster, with every channel above about 0.6. In contrast, the backdoored model’s activations are spread over a much wider range (about 0.2–1.7), and a large number of channels falls below the lowest activation of any channel in the benign model. These finding reveals that, although clean samples do not active the channel with backdoor embedded, they rather remain dormant on clean samples by showing a wider spread and lower activation magnitude. Therefore, we can conclude that backdoor embedded in the model can always result in activation differences compared benign models no matter to the presence of the trigger in the input. In this revised manuscript, we have created a new section to included the additional results and explanation to our channel activation inconsistency in Appendix B line 842.

Moreover, even under the scenario where malicious clients submit the CAS statistics computed purely on clean samples, the TF-IDF can still effectively capture the abnormal pattern from them. Since the activations have a wider spread across different channels, the TF value will also be spread, resulting abnormal CAS score distribution compared to benign ones. The following table shows the results when malicious clients submit the CAS statistics computed clean samples:

Table R3-2: Performance of CAS when malicious clients compute CAS statistics on clean samples

DatasetAttackAccASRTPRFPR
CIFAR-10BadPFL95.13 ± 0.3213.02 ± 2.8793.30 ± 1.928.03 ± 3.09
CIFAR-10IBA95.12 ± 0.3210.16 ± 1.0097.40 ± 1.854.83 ± 2.05
CIFAR-10DBA95.07 ± 0.3310.05 ± 0.1399.40 ± 1.083.92 ± 1.86
CIFAR-10Neurotoxin95.14 ± 0.3410.02 ± 0.1399.20 ± 0.454.30 ± 1.07
CIFAR-100BadPFL78.56 ± 0.811.97 ± 1.1695.40 ± 1.711.30 ± 1.06
CIFAR-100IBA78.49 ± 0.591.23 ± 0.1999.70 ± 0.450.13 ± 0.15
CIFAR-100DBA78.38 ± 0.761.11 ± 0.0399.80 ± 0.270.18 ± 0.11
CIFAR-100Neurotoxin78.39 ± 0.731.04 ± 0.0399.80 ± 0.270.93 ± 0.39

As shown in Table R1-3, CAS stays effective under this adaptive attack. On both datasets it detects at least 93% of malicious clients and keeps ASR close to random guess. This shows that CAS does not depend on whether the class-wise activation statistics are computed on triggered or clean samples. The backdoor embedded in the model alone is enough to produce detectable anomalies in the CAS scores. In the revised manuscript, we have incorporated this adaptive attacker scenario in section 3.1 line 467.

AuthorsOpenReview ↗

Response to Reviewer oWNB (3/3)

Response to Critical Weakness 3:

We thank the reviewer for this comment. As demonstrated above, benign clients' channels are lower bounded given the premise that they capture features from the same input distribution. However, channels with backdoor embedded are inevitably deviated from the normal channel by showing abnormal dormancy, resulting lower activation entropy regardless to data distribution. In this case, the cluster grouped by malicious clients will often have a lower center norm due to their lower activation entropy. Moreover, since malicious clients often have fixed malicious objective (e.g., same target class), they inevitably tend to have aligned lower entropy behaviors towards the target class. In contrast, the low entropy behaviors of all benign clients often do not bind to a particular class. In such case, our proposed clustering mechanism can effectively capture the CAS abnormal pattern among all malicious clients and exclude them from aggregation. To support our claim, we provide the experimental results showing the superior defense performance of CAS under various degree of data heterogeneity:

Table R3-3: Performance of CAS under different degrees of data heterogeneity (Dirichlet α\alpha) on CIFAR-100

α\alphaAttackACCASR ↓TPR ↑FPR ↓
0.1DBA75.430.94100.000.63
0.1IBA75.720.82100.000.00
0.1BadPFL76.011.1496.501.00
0.1Neurotoxin75.700.96100.000.37
0.3DBA76.621.14100.000.13
0.3IBA77.411.0398.000.25
0.3BadPFL77.351.6693.501.38
0.3Neurotoxin76.871.07100.000.13
0.5DBA77.761.14100.000.00
0.5IBA77.620.9599.500.13
0.5BadPFL77.921.3397.500.13
0.5Neurotoxin77.711.13100.000.25

Experiment logs for Table R3-1 to R3-3 can be found in here.

Once again, we sincerely thank Reviewer oWNB for his/her time and the constructive and insightful comments. We look forward to any further discussion with you.

References:

[1] Li et al., "BackdoorIndicator: Leveraging OOD data for proactive backdoor detection in federated learning", USENIX Security 2024.

[2] Kadir et al., "Backdoor Attacks Rerouted: Batch Normalization as a Sink for Adversarial Signals", NeurIPS 2026.

[3] Huang et al., "Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training", NeurIPS 2023.

[4] Huang et al., "Parameter Disparities Dissection for Backdoor Defense in Heterogeneous Federated Learning", NeurIPS 2024.

[5] Yan et al., "FedRACE: A Hierarchical and Statistical Framework for Robust Federated Learning", NeurIPS 2025

Reviewer_A2rdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper address the problem of backdoor defense for FL with pretrained model. This paper found that the model updates from benign and malicious clients become increasingly similar when the global model is initialized with pre-trained weights. To tackle this problem, this paper test the channel activation behavior, and observing that the channel activiations of benign models relatively consistent, whereas backdoored models shows anomaly class activiation patterns.

Based on the above observations, this paper propose CAS, which applies the TF-IDF to meaure the contribution of each channels to specific classes, calculate the degree of activiation anomaly by using information entropy, and use the clustering based on Wasserstein distance to identfiy the malicious client from aggregation.

Critical strengths
  1. I think the key strength of this paper is to reveal that when federated learning is initialized with a pre-trained model, the model updates from benign clients and malicious clients become more similar. This significantly decrease the effectiveness of existing backdoor defense methods that rely on differences in model updates.
  2. This paper did a relative comprehansive experiments with different pre-training approaches, non-IID data distributions, varying proportions of malicious clients, and adaptive attacks.
Critical weaknesses
  1. This paper apply TF-IDF to construct Channel Activation Distribution. However, for any given channel, the IDF value remains constant across different classes. Therefore, it will be canceled out during the normalization process. This implies that the resulting Channel Abnormality Score may depend solely on the TF component, rendering the IDF irrelevant. Furthermore, the normalization step could lead to undefined values ​​in cases where the IDF is zero. This issue directly impacts the paper’s primary technical contribution, as the paper position TF-IDF as a key design element of the CAS which claim that the current formulation fails to support.
  2. This paper claim that channel activations in benign clients tend to be more consistent, whereas those in malicious clients exhibit more pronounced, class-specific activation patterns where a distinction that can be captured using entropy. However, certain channels within a benign model may also respond strongly only to specific classes or features, particularly under non-IID data distributions. Therefore, low entropy does not necessarily indicate the presence of a backdoor.
Questions for authors
  1. Could the authors explain the actual role of TF-IDF in the final detection score? Does the actual implementation align with the formula? Would the same results be obtained if the IDF component were removed? Especially, In Equation (6) the IDF is constant for all categories within the same channel. Thus, it cancels out during the normalization process. This implies that the final Channel Abnormality Score may depend solely on the TF and independent of the IDF. Furthermore, when the IDF is zero, the normalization may become undefined.
  2. This paper employ the one-dimensional Wasserstein-1 distance for clustering; however, the cluster centers in Equation (8) are calculated by taking the arithmetic mean after sorting the data. In the context of the Wasserstein-1 distance, the optimal cluster center should generally be determined by the coordinate-wise median, rather than the arithmetic mean. Could the authors explain the rationale behind this choice of center update method? Does this implementation effectively optimize the Wasserstein-1 clustering objective as defined in the paper? Furthermore, does this approach impact the clustering results obtained by the clients?
  3. This paper claims that channel activations in a normal model tend to be relatively consistent, whereas a backdoored model performs distinct. However, I think such behavior could also occur in a normal model, for example, certain channels might naturally respond more strongly to specific classes or visual features, especially in non-IID scenarios. Could the authors elaborate on why low entropy necessarily implies the presence of a backdoor? In addition, how can one distinguish between normal class-specific features and the anomalous activations triggered by a backdoor?
  4. CAS divides clients into two clusters, positing that those with a larger Center Norm are the benign ones. I am not entirely convinced that this assumption always holds. For instance, in a non-IID setting, the activation distributions of different benign clients can vary significantly. Therefore, malicious clients employing different types of attacks do not necessarily share similar characteristics. Could the authors explain why partitioning clients into only two clusters is sufficient? Furthermore, why does a larger Center Norm imply that a client is benign?
7 comments
AuthorsOpenReview ↗

Response to Reviewer A2rd (1/3)

We thank reviewer A2rd for the feedback and valuable suggestions. Below we provide responses to each weakness and question. Please feel free to reach out if further clarification is required.

Response to Major Weakness 1:

We thank the reviewer for pointing this out. The reviewer is correct that, as printed, the IDF term cancels in Eq. 6. We sincerely apologize that this is caused by a typo in the denominator of Eq. 6: the summation should run over the CoutlC^l_{\text{out}} channels of layer ll, not over the NN classes. The corrected normalization is

Pkl(n,j)=TF-IDFkn,l,j∑j′=1CoutlTF-IDFkn,l,j′+ϵ.\mathcal{P}_k^l(n,j) = \frac{\text{TF-IDF}_k^{n,l,j}}{\sum_{j'=1}^{C^l_{\text{out}}} \text{TF-IDF}_k^{n,l,j'}+\epsilon}.

In the erroneous version, IDFl,j\text{IDF}^{l,j} is identical for every class nn, so it is a common factor of the numerator and the denominator and cancels. In the corrected version, IDFl,j\text{IDF}^{l,j} varies across the channels being summed, so it does not cancel. Instead, it reweights the class-wise channel distribution: channels that are highly active for only a few classes (high IDF) receive more probability mass than channels shared by many classes (low IDF). This is precisely the class-specific concentration that CAS is designed to expose, and it is why the IDF term affects the abnormality score zkl,nz_k^{l,n}. We have made corresponding corrections in the revised manuscript in section 2.2 line 268. In addition, we have re-run all experiments impacted by this mistake and updated the results in the manuscript. The logs can be found here.

Response to Major Weakness 2:

We thank the reviewer for this insightful comment. We agree with the reviewer that pronounced, class-specific activation patterns can also occur in channels that respond strongly only to specific classes or features. However, we would like to justify that, even though these channels activate in a class-specific manner, they are not completely dormant to the reset of classes as class-specific features are also part of the nature feature from the same input distribution [1]. In contrast, backdoor shortcut induced by malicious triggers is not something nature from the input feature distribution. Under such case, channels activate strongly to the backdoor trigger can largely remain dormant to clean samples, resulting abnormal activation pattern overall. Therefore, although Non-IID data distribution can vary the channel activation distribution of benign clients, their magnitudes are still lower bounded given the premise that they come from the same input feature distribution. However, since backdoored samples can be essentially seen as out-of-distribution samples with the respect to the benign inputs [2] [3], channels with backdoor embedded are inevitably deviated from the normal channel by showing abnormal dormancy regardless to data distribution. To support our claim, we provide the experimental results showing the superior defense performance of CAS under severe data heterogeneity:

Table R2-1: Performance of CAS under different degrees of data heterogeneity (Dirichlet α\alpha) on CIFAR-100

α\alphaAttackACCASR ↓TPR ↑FPR ↓
0.1DBA75.430.94100.000.63
0.1IBA75.720.82100.000.00
0.1BadPFL76.011.1496.501.00
0.1Neurotoxin75.700.96100.000.37
0.3DBA76.621.14100.000.13
0.3IBA77.411.0398.000.25
0.3BadPFL77.351.6693.501.38
0.3Neurotoxin76.871.07100.000.13
0.5DBA77.761.14100.000.00
0.5IBA77.620.9599.500.13
0.5BadPFL77.921.3397.500.13
0.5Neurotoxin77.711.13100.000.25
AuthorsOpenReview ↗

Response to Reviewer A2rd (2/3)

Response to Question 1:

  • "Could the authors explain the actual role of TF-IDF in the final detection score? Does the actual implementation align with the formula? Would the same results be obtained if the IDF component were removed?"

TF-IDF serves as the centre role for our proposed malicious detection method. It essentially turns the observation of channel activation inconsistency into a measurable malicious detection signal. In our updated codes, all formulas align with the implementation (particular for TF-IDF calculation which can be found in ./defenses/TF_IDF_calculator.py). Moreover, in the following table, we show the performance of CAS with and without IDF term incorporated. With the corrected Eq.(6) with IDF correctly incorporated, it pushes the performance of CAS even better. Besides, to avoid the division by zero error, we have added a small number ϵ\epsilon in the denominator to stabilize the calculation. In the following table, it can be observed that correctly incorporating IDF term results in substantial performance gain in terms of ASR reduction and ACC preservation.

Table R2-2: Performance of CAS without and with the IDF term (averaged over BadPFL, IBA, DBA and Neurotoxin attacks)

DatasetIDFACC ↑ASR ↓TPR ↑FPR ↓
CIFAR-100✗78.961.7495.621.29
CIFAR-100✓79.501.1499.150.27
CIFAR-10✗94.9111.8094.385.16
CIFAR-10✓95.4710.4697.355.65
SVHN✗97.0914.9696.880.12
SVHN✓97.1512.2398.620.09

Response to Question 2:

  • "Could the authors explain the rationale behind this choice of center update method? Does this implementation effectively optimize the Wasserstein-1 clustering objective as defined in the paper?"

We thanks the reviewer for this question. The element-wise median often provides a robust estimate while suppressing outlier effects. However, the core of our CAS defense is to catch any potential outlier that demonstrate abnormal activation pattern. In this case, taking the mean after sorting can be a better choice for finding the new cluster center. In our implementation, we calculate the coordinate-wise mean within each newly assigned cluster center and apply assign clients using scipy.stats.wasserstein_distance.

  • "Does this approach impact the clustering results obtained by the clients?"

The following table compares the defense performance between using mean and median as the center update metric:

Table R2-3: Performance of CAS with median versus mean as the cluster-center update on CIFAR-100

AttackACC ↑ (median / mean)ASR ↓ (median / mean)TPR ↑ (median / mean)FPR ↓ (median / mean)
BadPFL78.55 / 78.391.14 / 1.1396.50 / 97.001.25 / 1.00
IBA78.63 / 78.771.18 / 0.9799.00 / 99.500.00 / 0.00
DBA78.68 / 78.701.06 / 1.09100.00 / 100.000.00 / 0.00
Neurotoxin78.72 / 78.691.05 / 1.0499.50 / 100.000.25 / 0.00
Average78.65 / 78.641.11 / 1.0698.75 / 99.130.38 / 0.25

As shown, using mean can generally result in marginal improvements compared to that of using median.

Response to Question 3:

  • "Could the authors elaborate on why low entropy necessarily implies the presence of a backdoor?"

We appreciate the reviewer for brining this question. As revealed in our reply to weakness 2, malicious channels are generally more dormant to the nature feature from the input distribution, resulting low activation entropy overall. More details can be found in our reply to the critical weakness 2.

  • " In addition, how can one distinguish between normal class-specific features and the anomalous activations triggered by a backdoor?"

Motivated by this property, CAS introduces the IDF term to balance the suspiciousness of a class-specific channel by taking the activation on other classes into account.

AuthorsOpenReview ↗

Response to Reviewer A2rd (3/3)

Response to Question 4:

  • "Why does a larger Center Norm imply that a client is benign?"

We thank the reviewer for this question. As demonstrated above, benign clients' channels are lower bounded given the premise that they capture features from the same input distribution. However, channels with backdoor embedded are inevitably deviated from the normal channel by showing abnormal dormancy, resulting lower activation entropy regardless to data distribution. In this case, the cluster grouped by malicious clients will often have a lower center norm due to their lower activation entropy. To support our claim, we plot the relationship between the two cluster in terms of their center norm and the number of malicious client contained. The link to the plot is here. The experiment is conducted under a non-IID degree of Dirichlet 0.1 and the attackers are free to choose between four attacks {Neurotoxin, BadPFL, IBA, BDA}. As shown, the larger size cluster, with a larger center norm, does not contain any malicious clients in most of time. In contrast, the smaller size cluster, with a smaller center norm, often contains all malicious clients in most of time. In the revised manuscript, we have included the plot showing the center norm difference as rounds progress in Section 3.1 line 451. Meanwhile, we also show the superior defense performance of CAS under this mutli-attacker scenario in Appendix G.

  • "Could the authors explain why partitioning clients into only two clusters is sufficient? "

Since 2-means partitions along the dominant direction of variation (particularly for the anomalous signal [4]) in the sample space [5], clustering clients into two group enables K-means to more focus on the abnormal activation pattern of backdoor embedded channels in malicious models. A larger K can potentially shift its focus towards other factors (e.g., data/model/domain heterogeneity). As shown in the following table, as we increase the number of intend clusters and drop the one with the lowest center norm, the defense performances gradually degrades:

Table R2-4: Defense performance of CAS with different numbers of clusters KK (CIFAR-100, Attack: DBA)

KKACC ↑ASR ↓TPR ↑FPR ↓
2 (default)80.691.05100.000.00
379.014.6487.001.25
478.8530.7472.500.13

Experiment logs for Table R2-1 to R2-4 can be found in here.

Once again, we sincerely thank Reviewer A2rd for the constructive and insightful comments. We look forward to any further discussion.

References:

[1] Liu et al. "Fine-pruning: Defending against backdooring attacks on deep neural networks." International symposium on research in attacks, intrusions, and defenses. Cham: Springer International Publishing, 2018.

[2] Li et al., "BackdoorIndicator: Leveraging OOD data for proactive backdoor detection in federated learning", USENIX Security 2024.

[3] Sui et al., "Backdoor Detection in Federated Learning With Feature Map: A Multi-Task Learning Perspective", IEEE TIFS 2026.

[4] Chen et al. "Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering", arXiv 2018.

[5] Ding et al. "K-means clustering via principal component analysis", ICML 2004.

Reviewer_U2XrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper examines channel activation inconsistency to detect backdoors in federated learning (FL) when pretrained backbones are used, where standard update-based defenses are less effective. The authors introduce CAS, a method that computes per-channel abnormality scores from class-wise activation statistics uploaded by clients. CAS then uses Wasserstein-distance-based 2-means clustering to split clients into two groups and aggregates the cluster with the larger center norm. Experiments on CIFAR-10/100, SVHN, and Tiny-ImageNet include four attacks and six defense baselines (AlignIns, MARS, FDCR, Lockdown, Trimmed-Mean, Multi-Krum), along with undefended FedAvg.

Critical strengths
  • The paper provides a convincing diagnosis of why update-similarity defenses weaken with pretrained backbones. The layer-replacement study (shown in Fig. 1c–d and Appendix A) clearly links pretraining to a smaller benign–malicious update gap and to higher ASR under existing defenses.
  • CAS leverages class-conditional channel activations as a complementary detection signal with low communication overhead. Experiments across four datasets, four attacks, and multiple baselines show consistent ASR reduction and low FPR in most of the evaluated FL settings.
Critical weaknesses

Major Weaknesses

  1. In this work, the CAS relies on client-reported activation statistics that malicious clients can easily forge. If checked carefully, the proposed authentication (Eq. 13) only checks whether the per-channel sum over classes matches the BN running mean, which clients also control. This leaves the class-wise pattern that CAS scores unconstrained. I do not see any experiment that tests a statistic-forging attacker, and the main results (Tables 1–3) are obtained without authentication. Also, the adaptive attack (Eq. 12) overlooks this simple evasion route, leaving the core security assumption unvalidated.
  2. Because CAS always splits clients into two clusters and discards one, it inevitably removes benign clients when few or no attackers participate. Under intermittent attacks (Table 9, r = 20%), its FPR reaches 16–23%. For Neurotoxin, this exceeds the FPR of MARS and FDCR, contradicting the claim of "substantially lower FPR". I am expecting to see the no-attacker setting and its effect on model utility.

Minor Weaknesses

  • The IDF term cancels out in Eq. 6, so the TF-IDF framing does not affect the score.
  • Averaging ASR across all rounds may inflate the apparent defense effect.
  • Uploading class-wise statistics leaks each client's label distribution, which weakens the privacy argument.
  • Without pretraining, CAS has the lowest TPR among the compared defenses (Table 3).
  • Related activation-based backdoor defenses, such as Fine-Pruning, Activation Clustering, DeepSight, and FLAME, are not discussed.
  • I found some numerical inconsistencies across Tables 1, 2, 3, 7, 8, and 13. E.g., Trimmed-Mean averages in Table 1 do not match their own rows. For the same CAS setting, Table 2 reports an FPR of 7.08 while Table 13 reports 1.08. One Multi-Krum row in Table 8 duplicates values from Table 4 and implies an infeasible number of removed clients.
Questions for authors
  1. Can CAS detect an attacker who uploads forged statistics while keeping the BN buffer consistent so that Eq. 13 passes? Examples include statistics from a clean surrogate model, or statistics computed with poisoned samples assigned to their source labels. Does the server verify the uploaded statistics against anything the client does not control?
  2. What are CAS's FPR and accuracy when no attackers are present, or when the number of attackers varies per round? How do you reconcile the 16–23% FPR in Table 9 with the claim of "substantially lower FPR"?
  3. Could you correct or explain the inconsistencies in Tables 1, 2/13, 3, 7, and 8, and release per-round logs?
  4. Does the IDF term cancel in Eq. 6? What is the final-round ASR, as opposed to the round-averaged value?
7 comments
AuthorsOpenReview ↗

Response to Reviewer U2Xr (1/4)

We thank reviewer U2Xr for the feedback and valuable suggestions. Below we provide responses to each weakness and question. Please feel free to reach out if further clarification is required.

Response to Major Weakness 1:

We thank the reviewer for pointing this out. We agree with the reviewer that an adaptive attacker can bypass the BN statistics authentication by simultaneously manipulating the BN layers. However, as suggested in [1] [2], BN layers often encode a substantial amount of backdoor information to enable effective triggered behavior. That is, the training data of a backdoor task can be equivalently seen as a form of out-of-distribution data relative to that of the main task. Due to change of the training data, the estimated mean μ\mu and variance σ\sigma after the backdoor task gradually deviate from those of clean task. Thus, in order to bypass the BN authentication, attackers must confine to upload the BN statistics that are not longer capable to maintain the backdoor effect. In other words, the proposed BN authentication aims to drive the adaptive attackers into a dilemma where they either get caught by the authentication or suffer from severe backdoor performance degradation. To support our claim, we consider three forms of adaptive attackers:

  • Type-1: the attacker additionally trains a benign surrogate and uploads the CAS statistics of this model.
  • Type-2: the attacker additionally trains a benign surrogate and replaces the BN layers of the backdoored model with that of the benign one.
  • Type-3: the attacker additionally trains a benign surrogate and rescales CAS statistics of the benign model by the ratio between the backdoored and the benign BN running means ensure that it can sum to the backdoored model’s BN running mean, and uploads the rescaled CAS statistics.

The following table demonstrates the defense performance of CAS with the proposed BN authentication under the aforementioned adaptive attack scenarios:

Table R1-1: Performance of BN authentication against adaptive attackers in CIFAR10

AttackAdaptive TypeAccASRTPRFPR
DBAno adaptive95.2410.0199.500.50
DBA195.0710.11100.00.32
DBA293.1610.790.000.00
DBA395.4010.1999.5011.00
IBAno adaptive95.4311.7399.006.38
IBA195.1614.13100.004.30
IBA293.6612.630.000.00
IBA395.748.93100.000.33

As shown, CAS with BN authentication enabled provides effective defense service under all adaptive attack scenarios, minimizing the ASR rate down to random guess level. Under type-2, attacker bypass the CAS defense, but, due to the change of BN, the attack performance suffers from significant drop. Other FL defenses that additionally requires local clients to upload defensive statistics (e.g., Lockdown[3], FDCR[4], FEDRACE[5]) often do not offer a practice solution for the aforementioned adaptive attackers or simply assuming that the local gradients are unbreakable. In contrast, CAS naturally provides the guaranteed robustness under adaptive attackers. In the revised manuscript, we have incorporated aforementioned justification and results in Appendix F line 1066.

Response to Major Weakness 2:

We thank the reviewer for rasing this concern. We agree that the design of directly splitting clients into two clusters can inevitably result in increased FPR under the intermittent attack. That is, when there are no attackers in the FL system, blindly discarding one cluster according to the CAS pipeline might slow down the global model’s convergence or even reduce its accuracy. To address this issue, we propose to employ a "clustering brake" that stops cluster discarding if the the distance between the two clusters does not exceed a threshold ν\nu=10. Given that an FL system with only attackers is meaningless, we assume that all local models are benign when the cluster distance is low. Below, we provide the updated results under unsynchronized Attackers (rr=20%) and the scenario where no attackers are participated:

Table R1-2: Effectiveness of clustering brake under CIFAR-100

AttackScenarioAccASRFPR
BadPFLUnsynchronized (rr=20%)80.892.430.32
BadPFLNo attackers81.050.960.30
IBAUnsynchronized (rr=20%)81.041.130.21
IBANo attackers80.871.010.10
DBAUnsynchronized (rr=20%)80.921.170.00
DBANo attackers81.161.100.10
NeurotoxinUnsynchronized (rr=20%)81.171.090.32
NeurotoxinNo attackers80.871.040.10

In this case, we could say that when the "clustering brake" employed, the FPR is substantially lowered. In this revised manuscript, we have included this new component in Section 2.3 and updated the results for unsynchronized attackers in Appendix F line 1242.

AuthorsOpenReview ↗

Response to Reviewer U2Xr (2/4)

Response to Minor Weakness 1:

We thank the reviewer for pointing this out. The reviewer is correct that, as printed, the IDF term cancels in Eq. 6. We sincerely apologize that this is caused by a typo in the denominator of Eq. 6: the summation should run over the CoutlC^l_{\text{out}} channels of layer ll, not over the NN classes. The corrected normalization is

Pkl(n,j)=TF-IDFkn,l,j∑j′=1CoutlTF-IDFkn,l,j′+ϵ.\mathcal{P}_k^l(n,j) = \frac{\text{TF-IDF}_k^{n,l,j}}{\sum_{j'=1}^{C^l_{\text{out}}} \text{TF-IDF}_k^{n,l,j'}+\epsilon}.

In the erroneous version, IDFl,j\text{IDF}^{l,j} is identical for every class nn, so it is a common factor of the numerator and the denominator and cancels. In the corrected version, IDFl,j\text{IDF}^{l,j} varies across the channels being summed, so it does not cancel. Instead, it reweights the class-wise channel distribution: channels that are highly active for only a few classes (high IDF) receive more probability mass than channels shared by many classes (low IDF). This is precisely the class-specific concentration that CAS is designed to expose, and it is why the IDF term affects the abnormality score zkl,nz_k^{l,n}. We have made corresponding corrections in the revised manuscript in section 2.2 line 268. In addition, we have re-run all experiments impacted by this mistake and updated the results in the manuscript. Logs can be found in here.

Response to Minor Weakness 2:

We agree with the reviewer that averaging ASR across all rounds may inflate the defense effect. In the following, we report the highest ASR (%) reached by the global model in any round of training (peak ASR) under the cross-silo setting:

Table R1-3: CIFAR-10

DefenseBadPFLIBADBANeurotoxinAvg
FedAvg97.8399.9888.2797.4695.89
Trimmed-Mean92.4494.1095.0198.3594.97
Multi-Krum88.2590.5092.7598.5992.52
Lockdown100.00100.0099.87100.0099.97
FDCR93.6099.2097.1599.5297.37
MARS85.8186.3891.9998.6290.70
AlignIns99.8389.6292.2097.8794.88
CAS23.3311.8910.3710.3914.00

Table R1-4: CIFAR-100

DefenseBadPFLIBADBANeurotoxinAvg
FedAvg87.1597.7970.4897.3088.18
Trimmed-Mean69.9298.6691.6696.5389.19
Multi-Krum97.0297.5386.4397.7094.67
Lockdown49.8918.7229.673.3925.42
FDCR96.1999.9684.2997.7094.53
MARS73.0081.7475.8794.2481.21
AlignIns70.0082.1190.3597.9885.11
CAS1.991.411.311.201.48

As shown in Table R1-4 and R1-5, CAS constantly achieves best ASR resistance compared to baseline defenses.

Response to Minor Weakness 3:

We thank the reviewer for raising the privacy concerns. CAS statistics are calculated based on the channel statistics computed from spatially average-pooled activation magnitudes. These statistics are substantially more compressed than the underlying representations and are similar to the channel-wise summary statistics used by normalization mechanisms, such as means and variances. which is commonly adopted in most federated learning setups. In this case, the marginal privacy leakage exposed by CAS is equivalent to these normalization mechanisms. However, we agree that the equivalence between CAS and batch normalization statistics does not fundamentally lead to the degree of safety as rendered in our paper. Thus, we will re-position the privacy aspect of CAS more as a benefit (e.g., less sensitive compared to test accuracies, local prototypes, or other forms of feature representations) rather than rendering it as a part of the main contribution. In the revised manuscript, we have added the following clarification in Sec. 2.2 line 273: Compared to existing updates discrepancy-based defenses that exploit the test accuracies or the raw feature representations of local clients, the TF-IDF statistics of each client does not expose direct connection to the local private data.

AuthorsOpenReview ↗

Response to Reviewer U2Xr (3/4)

Response to Minor Weakness 4:

We thank the reviewer for this comment. We acknowledge that CAS does not guarantee state-of-the-art defense performance when the global model is trained from scratch (TS), as CAS is primarily designed for federated learning with a pretrained global model. The reason lies in how pretraining shapes channel activations. As shown in Figure 2, compared with its TS counterpart, a pretrained global model has a much larger fraction of channels whose activations on backdoored samples are near zero, while their activations on clean samples are considerably stronger. In other words, pretraining makes the channel activations of malicious clients far more inconsistent across classes, and this is exactly the signal CAS relies on. Without pretraining, this discrepancy is weaker, which explains the moderate performance drop of CAS in the TS setting (Table 3).We believe this dependence does not limit the practicality of CAS. This is because whether to initialize the global model from a pretrained backbone is entirely the server’s decision, so the server can always meet the condition under which CAS is most effective.

Response to Minor Weakness 5:

We thank the reviewer for rasing the discussion. In the following, we provide detailed discussions regarding the related activation-based backdoor defenses:

  • Fine-Pruning: As a post-training defense, Fine-Pruning assumes that backdoor channels generally remain dormant on clean validation data, and it therefore prunes these channels to repair the backdoored model. In contrast, CAS extends this premise (Section 2.2) and uncovers that channel activations can serve not only as a criterion for neuron pruning but also as an effective signal for detecting malicious clients in federated learning. Specifically, CAS measures how strongly each channel’s activation is concentrated on individual classes, compares these class-wise activation patterns across clients, and filters out the clients whose patterns are abnormal.

  • Activation Clustering: Activation Clustering is motivated by the finding that, within the same class, the activations of poisoned and clean samples form two distinct clusters. This allows the defender to identify the poisoned samples and retrain a clean model, but it requires centralized access to the training data. In contrast, CAS reveals that benign and malicious models exhibit different activation patterns across classes. Malicious models concentrate a channel’s activation on the target class, whereas benign models activate it more consistently across classes. Because these patterns can be summarized as class-wise running means computed locally on each client, CAS turns this observation into an effective signal for detecting malicious clients in federated learning, without the server accessing any client data.

  • DeepSight: DeepSight fundenmatally does not explores the role of activation pattern in malicious detection and rely on updates magnitudes of the classification head to assess whether certain classes are perturbued by backdoor. Similar to other gradient-based detection methods (e.g., Trimmed-Mean, AlignIns), with a pretrained initial global model, output-layer update magnitudes may shrink and become similar across clients.

  • FLAME: FLAME is another gradient-based defense that works entirely in parameter space (cosine distances, L2 norms, noise). With a pretrained global model, benign and malicious updates are small and close together as shown in Figure 1 in Section 1.

We have included the discussion to related activation-based defenses in Section. 4 line 519. Since Fine-tuning and Activation Clustering requires direct access to the local training data, which violates the privacy constraint of federated learning, we instead compare the defense performance of DeepSight and Flame against CAS as following

Table R1-5: Peak ASR / Mean ASR (%) of DeepSight, FLAME, and CAS

DatasetDefenseBadPFLIBADBANeurotoxinAvg
CIFAR-10DeepSight43.36 / 21.8821.27 / 13.2711.20 / 10.3038.95 / 15.8628.70 / 15.33
CIFAR-10FLAME99.89 / 93.6398.35 / 76.7396.20 / 71.8998.24 / 88.1898.17 / 82.61
CIFAR-10CAS23.33 / 11.5211.89 / 10.3910.37 / 9.9610.39 / 9.9914.00 / 10.46
CIFAR-100DeepSight93.24 / 76.3997.25 / 82.4683.11 / 54.1797.58 / 78.2292.80 / 72.81
CIFAR-100FLAME93.19 / 82.2399.33 / 86.8694.94 / 79.0996.73 / 84.7396.05 / 83.23
CIFAR-100CAS1.99 / 1.291.41 / 1.111.31 / 1.111.20 / 1.041.48 / 1.14

As shown in Table R1-5, CAS significantly outperforms the two baselines defenses in terms of both peak ASK and average ASR.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers (U2Xr, A2rd, and oWNB) for her/his time reviewing our paper and providing insightful comments. In this work, we present the first systematic investigation exposing the fundamental vulnerability of existing federated backdoor defenses when confronted with pretrained global model. To address this limitation, we introduce CAS, a pioneering framework that investigate the client-discrepancy from the perspective of channel activation and detects attackers by analyzing the activation pattern among clients. All three reviewers generally found our paper easy to follow and appreciate the identified problem as concerning and interesting. oWNB also noted that our proposed solution is novel as it provides a new perspective for backdoor detection in federated learning.

The main concerns generally fall into four aspects:

  • (1) Adaptive attackers (U2Xr, oWNB): Since clients compute and upload CAS statistics locally, a BN-based authentication protocol is proposed to prevent adaptive attackers from manipulating these statistics to bypass the server-side detection. However, both reviewers point out that attacker can manipulate model updates should also be able to manipulate the accompanying activation statistics, rendering the authentication ineffective.

  • (2) Channel Activation Inconsistency (A2rd, oWNB): The observation depicted in section 2.2 demonstrates the activation pattern of backdoored models using poisoned samples whereas shows that of benign ones using clean samples. Thus, it is unclear that the observed activation differences arise from the backdoor embedded in the model or merely from the presence of the trigger in the input. Moreover, under highly non-IID scenario, benign model could also demonstrate low activation entropy since a benign model may also respond strongly only to specific classes or features. Thus, it is unsure that low entropy necessarily indicate the presence of a backdoor.

  • (3) Clustering Mechanism (U2Xr): The proposed clustering mechanism drops the cluster with a smaller center norm. However, it is unclear that how does this mechanism work when there is no attackers.

  • (4) Cancellation of the IDF Component (U2Xr, A2rd): Both reviewer noted that the IDF component can be canceled out under Eq. 6. Thus, it is unclear what is the role the IDF in CAS.

Regarding each concern, we summarize our response in the following:

  • (1): Existing studies have shown that BN layers often encode a substantial amount of backdoor information to enable effective triggered behavior. Thus, the proposed BN authentication aims to drive the adaptive attackers into a dilemma where they either get caught by the authentication or suffer from severe backdoor performance degradation.

  • (2): We unified the (poisoned) data used to produce activation patterns across all models in section 2.2 as suggested by oWNB. Then, we conduct additional experiments to investigate the activation behaviors of all clients on clean samples. The results revel that backdoor-active channels rather remain dormant on clean samples, indicating that backdoor embedded in the malicious model can always result in activation differences regardless of input data. In this case, although Non-IID data distribution can vary the channel activation distribution of benign clients, they are still lower bounded given the premise that they capture features from the same input feature distribution. In contrast, for channels with backdoor embedded, they are inevitably deviated from the normal channel by showing abnormal dormancy, leading to lower activation entropy.

  • (3): To incorporate the scenario when there is no attackers, we additionally introduce a "clustering brake" that stops cluster discarding if the the distance between the two clusters does not exceed the threshold ν\nu.

  • (4): The fact that the IDF term cancels in Eq. 6 is caused by an error. We have corrected it in the revised manuscript and update the results impacted by this mistake.

We are happy to clarify any remaining questions during the discussion period.

AuthorsOpenReview ↗

Summary of Responses and Revisions

Dear AC/SAC/PC,

We sincerely thank the ICLR reviewers, ACs, and PCs for their time and constructive feedback. Below, we would like to provide a final summary of our contributions, discussion, and revisions.

Contribution and Strength

In this work, we present the first systematic investigation exposing the fundamental vulnerability of existing federated backdoor defenses when confronted with pretrained global model. To address this limitation, we introduce CAS, a pioneering framework that investigate the client-discrepancy from the perspective of channel activation and detects attackers by analyzing the activation pattern among clients.

We are encouraged that the reviewers recognized the following strengths of our method:

  • Motivation. Reviewers found our empirical analysis regarding the failure of existing federated backdoor defenses under the pretrained setting to be insightful and convincing (U2Xr, A2rd).

  • Method Design. Reviewer (oWNB) marked our proposed new perspective of backdoor defense to be novel.

  • Experimental Strength. Reviewers found our empirical assessments in our work to be comprehensive (A2rd) and generalized (U2Xr,oWNB ), the proposed defense achieves show consistent backdoor effect minimization while resulting in low communication overhead (U2Xr).

  • Writing and Informativeness. All reviewers find our paper generally easy to follow and lead them to learn something useful.

Summary of Discussion

  • Reviewer U2Xr: 3 - Positive. Increased the score from 2 to 3, acknowledged that the additional experiments and clarifications addressed his/her concerns, and stated that "my concerns are mostly addressed now"

  • Reviewer A2rd: 2 - Acknowledged that Concern 1 is largely addressed. Raised three additional technical concerns and we provided further clarification to each of them. Given the limited remaining rebuttal time, we will reflect more throughout analysis and experiment results regarding these concerns in our revised manuscript. Notably, the second additional concern regarding adaptive attackers aligns to Reviewer oWNB’s Weakness 1, which oWNB acknowledged had been addressed.

  • Reviewer oWNB: 2 - Acknowledged that Concern 3 is largely addressed. Raised two additional technical concerns and we provided further clarification to each of them. Given the limited remaining rebuttal time, we will reflect more throughout analysis and experiment results regarding these concerns in our revised manuscript. Notably, the first additional concern regarding adaptive attackers overlaps with Reviewer U2Xr’s Major Weakness 1, which U2Xr acknowledged had been addressed.

Summary of Revision

  • Correcting the unintended cancellation of the IDF term in Eq. (6) and update impacted experimental results. [Reviewer U2Xr, A2rd]

  • Adding the privacy statement of CAS in Section 2.3 Line 270 and discussion on related channel-based backdoor defense in Section 4 line 519. [Reviewer U2Xr]

  • Incorporate the solution (Clustering Brake) for the non-attacker case in in Section 2.3 line 291 and updated the results for unsynchronized attackers in Appendix F line 1242. [Reviewer U2Xr]

  • Include the justification of solution to adaptive attackers and supporting experiment results in Appendix F Line 1066. [Reviewer U2Xr, oWNB]

  • Supporting the finding of channel activation inconsistency from the perspective of pure clean samples in Appendix B line 840 and Section 3.1 line 468. [Reviewer A2rd, oWNB]

  • Providing additional results on the relation between two clusters in Section 3.1 line 451[Reviewer A2rd]

Experimental logs during the rebuttal can be found here

We hope this summary assists in your evaluation. Thank you once again for your time and consideration.

Best regards,

Submission 6515 Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.