acceptodds
Under review as a conference paper at ICLR 2027

Eraser: Erasing Trigger-Induced Effects for Backdoor Defense in Federated Learning

Abstract

Federated Learning (FL) is highly vulnerable to backdoor attacks, where malicious clients implant trigger-dependent behaviors into the global model while preserving main-task performance. Existing defenses rely on malicious-client identification or training-time intervention, and often suffer robustness degradation under heterogeneous data, stealthy attacks, and diverse federated settings. We propose Eraser, a post-training backdoor defense for FL that erases trigger-induced effects before inference, preventing triggers from activating backdoor behaviors in a poisoned global model. Since potential triggers are unknown and difficult to remove directly from input samples, Eraser instead reconstructs sanitized samples from robust latent semantics. Motivated by our observation that standalone benign local models exhibit inherent robustness against backdoor triggers, Eraser uses their Featurizers to suppress trigger-related cues in the latent space, while a collaboratively trained Generator reconstructs sanitized samples from these representations. To improve cross-client consistency, we introduce fixed hyperspherical semantic anchors to align local feature spaces and stabilize generator training. Eraser is model-agnostic, compatible with most federated algorithms, and requires neither trigger knowledge nor malicious-client identification. Experiments on diverse attacks and heterogeneous settings show that, compared with existing defenses, Eraser achieves the best robustness while preserving main-task accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Backdoor Attack, Backdoor Defense

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ABjoOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Eraser, a post-training backdoor defense method for federated learning designed to mitigate the impact of backdoor triggers on model predictions without modifying the original federated training process or the potentially compromised global model. Unlike existing defense methods that rely on malicious client identification, robust aggregation, or model parameter repair, Eraser employs a semantic reconstruction strategy to eliminate trigger-related effects before inputs reach the global model. Specifically, each client independently trains a featurizer to extract semantic representations that are relatively insensitive to triggers, while feature alignment across clients is facilitated by fixed hyperspherical semantic anchors. Subsequently, clients collaboratively train a lightweight generator that utilizes classification supervision from the frozen global model to reconstruct latent semantic features into inputs suitable for processing by the global model. During inference, each client sequentially employs its local featurizer, the shared generator, and the frozen global model to generate predictions. The authors evaluated Eraser across various datasets, backdoor attacks, data heterogeneity scenarios, client scales, federated learning algorithms, and model architectures, while also investigating adaptive attacks, reconstruction behavior, and extensions to text and speech modalities. Experimental results demonstrate that the method significantly reduces the success rate of backdoor attacks while maintaining accuracy on benign tasks, all with minimal additional communication overhead.

Critical strengths

Strength 1: It proposes a post-training backdoor defense method that offers a degree of novelty. This paper introduces a defense strategy distinct from traditional approaches such as malicious client detection, robust aggregation, and model parameter repair. Eraser employs an independently trained local featurizer to extract semantic features relatively insensitive to backdoor triggers and reconstructs inputs via a shared generator; this enables defense implementation without altering the original federated training process or the poisoned global model.Shifting the defense mechanism to a pre-inference semantic reconstruction stage is an innovative design choice that also allows the method to be integrated with existing federated learning algorithms. I consider this approach to be the paper's most significant contribution. Strength 2: The experimental validation is comprehensive, demonstrating strong backdoor suppression capabilities across a wide range of settings. The authors not only conducted systematic comparisons across multiple datasets and typical backdoor attacks but also evaluated performance under various scenarios involving data heterogeneity, client scales, federated learning algorithms, and model architectures, as well as against adaptive attacks and in cross-modal extensions. Experimental results show that Eraser significantly reduces the success rate of backdoor attacks across diverse settings while maintaining high accuracy on benign tasks. The authors also provided ablation studies, visualizations of the reconstruction process, and an analysis of computational overhead, ensuring the completeness of the experimental validation. These results offer valuable empirical support for the method's effectiveness and potential applicability.

Critical weaknesses

Weaknesses 1: Existing evaluation metrics are insufficient to fully demonstrate the restoration of correct prediction capabilities for triggered samples. The paper primarily uses the backdoor failure rate to evaluate defense effectiveness. However, this metric merely indicates that triggered samples are not predicted as the attacker-specified target class; it fails to distinguish between samples correctly predicted as their true labels and those incorrectly predicted as other non-target classes. This issue is particularly critical for Eraser, as the method's core objective is to suppress the influence of the trigger through semantic reconstruction while preserving information relevant to the normal task. While a high backdoor failure rate demonstrates the suppression of the attack target's prediction, it does not directly prove that the reconstructed triggered samples have regained the correct classification result. Although the authors report favorable clean accuracy and high backdoor failure rates, these metrics address distinct evaluation goals—clean samples versus triggered samples—and cannot fully substitute for accuracy based on the true labels of triggered samples. Consequently, the current experimental validation of the core defense efficacy remains incomplete. Weaknesses 2: There is a lack of sufficient analysis regarding cross-client feature compatibility and applicability under conditions of local data scarcity. Eraser relies on a featurizer trained independently by each client to extract semantic features, while all clients share a common generator. Thus, the ability of different clients to maintain consistent feature representations is a crucial prerequisite for the effective operation of the shared reconstruction mechanism. The authors introduce hyperspherical semantic anchors and experimentally demonstrate a reduction in class-level feature discrepancies. However, class-level alignment does not necessarily imply consistency in the fine-grained semantic structures of samples within the same class across different clients. The paper currently lacks a direct analysis of this issue. Furthermore, Table 16 in Appendix J shows that, with the total training data volume held constant, Eraser's accuracy on the normal task drops significantly as the number of clients increases. Since the local featurizer does not participate in federated aggregation, this design may render performance heavily dependent on the data scale of individual clients. However, the paper does not adequately distinguish whether this accuracy decline stems primarily from insufficient training of the local featurizer or from performance fluctuations inherent to the original global model itself.

Questions for authors

Q1. Ability to recover correct labels for triggered samples The paper primarily uses the "backdoor failure rate" to measure defense effectiveness. However, this metric only reflects whether triggered samples avoid being predicted as the attacker-specified target class; it cannot distinguish between samples being correctly classified with their true labels and those being misclassified into other non-target classes. Consequently, a high backdoor failure rate does not necessarily imply that the model has recovered the ability to correctly predict triggered samples. Could the authors provide further results—specifically under typical backdoor attacks that preserve original sample semantics—regarding the classification accuracy of triggered samples relative to their original true labels after Eraser processing? This metric would help differentiate between the suppression of backdoor target predictions and the recovery of correct task predictions, thereby allowing for a more comprehensive evaluation of the defense's effectiveness. Q2. Compatibility of fine-grained feature representations across clients Eraser facilitates feature alignment across clients using fixed hyperspherical semantic anchors. The authors evaluate alignment effectiveness via feature differences between class means (Figure 4 and Appendix D) and demonstrate the shared Generator's generalization capability through experiments with new clients (Appendix J). However, class-level feature alignment does not necessarily guarantee sufficient consistency in the fine-grained feature distributions of samples belonging to the same class across different clients. Given that all clients utilize a shared Generator, could the authors further analyze intra-class feature distribution differences across clients, or explain why class-level semantic anchor alignment suffices to support stable reconstruction by the shared Generator? This would help further elucidate the method's feature compatibility and underlying mechanisms in data-heterogeneous environments. Q3. Trade-off between inference overhead and defense performance Tables 30 and 31 in Appendix S indicate that Eraser increases the computational cost of single-sample inference from 1.1158 GFLOPs to 2.2319 GFLOPs and inference latency from 3.1577 ms to 5.7539 ms—approximately 1.82 times that of the original global model. Although the additional communication cost introduced by this method is low, the inference phase still requires the execution of a local featurizer with a computational scale comparable to that of the global model; this may limit its practical deployment in resource-constrained or latency-sensitive scenarios. Given that the additional inference overhead stems primarily from the local featurizer, could the authors further analyze the relationship between the featurizer's computational scale and its backdoor defense performance? For instance, could employing a more lightweight feature extraction network significantly reduce inference overhead while maintaining comparable defense efficacy? Such an analysis would help clarify the trade-off Eraser makes between defense robustness and deployment efficiency. Q4. Sources of performance degradation under conditions of local data scarcity Table 16 in Appendix J shows that, while the total volume of training data remains constant, Eraser's accuracy on the main task drops from 84.95% to 74.18% as the number of clients increases from 100 to 300 (reducing the average number of training samples per client from approximately 449 to 149). The authors attribute this phenomenon to the reduction in training data available to the local featurizer; however, the current results do not sufficiently distinguish between the impact of diminished local feature-learning capabilities and changes in the underlying global model's performance. Could the authors provide the accuracy of the original global model on the main task—prior to applying Eraser—for the same client configurations, and compare these figures with the results obtained after applying Eraser? Such a comparison would help quantify the portion of performance degradation attributable to Eraser itself, thereby allowing for a more accurate assessment of the method's suitability in scenarios characterized by local data scarcity.

4 comments
AuthorsOpenReview ↗

Response (1/3)

We sincerely thank you for the insightful comments. Each one is answered below with new experiments. Unless otherwise specified, all experiments follow the default settings described in the manuscript.

Q1: Recovery of true labels for triggered samples

Response: Thank you for raising this important point. Eraser does more than divert triggered samples from the target class: it largely restores their true labels. To show this, we additionally report the recovery accuracy (RA), i.e., the accuracy of triggered samples with respect to their ground-truth labels. "Ideal" denotes a FedAVG global model trained without malicious clients and evaluated on the same triggered inputs, so its RA shows how much the trigger itself damages the input.

Table: RA (%) on CIFAR-10

BadNetDBASignalBlendedWaNetBppAttack1977Kelvin
Ideal78.2882.1742.4433.0385.1150.2167.6749.85
FedAVG0.000.000.000.0021.032.581.782.36
+Eraser86.2486.0250.9246.1686.8557.4072.1156.94

FedAVG's RA is near zero on seven of the eight attacks, whereas Eraser's RA exceeds even the Ideal model's on all eight, by 1.74% to 13.13%. On BadNet and DBA, Eraser's RA is within 1.7% of its clean accuracy (Table 2), so the trigger costs it almost no accuracy. RA is lower on Signal, Blended, BppAttack, and Kelvin because these are global triggers that alter the content of the entire image, which lowers the Ideal model's RA as well. Varying the blending ratio χ\chi of the Blended trigger in Eq. (20) confirms this: as the trigger grows stronger, the Ideal model's RA falls from 21.46% to 13.69%, while Eraser stays 10.8% to 19.0% above it.

Table: RA (%) under Blended triggers with blending ratio χ\chi

0.40.50.60.70.80.9
Ideal21.4620.7718.7615.5213.6713.69
FedAVG0.000.000.000.000.000.00
+Eraser40.5034.9534.3728.8229.3824.49

As requested, we also evaluate ISSBA [R1], whose invisible, sample-specific triggers preserve the original sample semantics. Eraser reaches an RA of 86.26%, within 2 percentage points of its clean accuracy and above the Ideal model (83.54%), while FedAVG drops to 0.13%.

Table: ISSBA on CIFAR-10 (%)

ISSBAABVRA
Ideal86.7798.4592.6183.54
FedAVG86.280.1343.210.13
+Eraser88.1799.1493.6686.26

We define RA in Appendix B and report these results, including ISSBA, in Appendix G (Table 12) of the revised manuscript. ISSBA is described in Appendix A.

[R1] Li et al. Invisible Backdoor Attack with Sample-Specific Triggers. ICCV 2021.

AuthorsOpenReview ↗

Response (2/3)

Q2: Fine-grained feature compatibility across clients

Response: Class-level anchors do more than align class means. LSTRL_{STR} in Eq. (4) is applied to every sample, pulling each feature toward its class anchor, and because all clients share the same anchors, the samples of each class are drawn toward the same region on every client.

We conduct this analysis with 20 clients on CIFAR-10, using the same data partitions across the two conditions. Each Featurizer undergoes 300 epochs of clean local training and is then frozen, followed by 50 communication rounds of shared Generator training. The two Generators start from identical weights and use the same frozen global model. Evaluation uses a common, class-balanced held-out set containing 100 images per class.

To verify this, we compare training with and without the anchors using the class-conditional MMD2\mathrm{MMD}^2 between clients, computed on raw features and, to remove class-mean offsets, on features centered by each client’s class mean. With anchors, the centered MMD2\mathrm{MMD}^2 falls by 33.60% (38.75% on classes seen by both clients), close to the 40.53% drop of the raw MMD2\mathrm{MMD}^2, so the alignment extends to the within-class distributions rather than stopping at the class means. To see whether this matters for the shared Generator, we measure the joint correct rate, the fraction of images that two clients both classify correctly when each passes the image through its own Featurizer and the shared Generator. With anchors it rises by 3.02%, so the same image is more often handled correctly no matter which client processes it.

Table: Effect of shared semantic anchors on cross-client feature compatibility under BadNet on CIFAR-10 (relative change for feature distances and absolute change in % for joint correct rate)

MetricRaw MMD2\mathrm{MMD}^2Centered MMD2\mathrm{MMD}^2Centered MMD2\mathrm{MMD}^2 (classes seen by both clients)Δ\Delta joint correct rate
Change with anchors−40.53%−33.60%−38.75%+3.02

We add this analysis to Appendix D.

AuthorsOpenReview ↗

Response (3/3)

Q3: Featurizer scale versus defense performance

Response: Lightweight Featurizers can substantially reduce computational cost while maintaining a similar B in this experiment, although the latency benefit depends on the architecture. We replaced the default ResNet-18 Featurizer with five lighter networks, keeping the Generator and the global model unchanged.

Table: Lightweight Featurizers under BadNet on CIFAR-10 (differences from the default ResNet-18 Featurizer)

Δ\DeltaB (%)Δ\DeltaGFLOPsΔ\Deltalatency (ms)
CNN+0.51-1.1052-0.8474
MobileNet-V2 ×0.5+0.43-1.0939+1.2904
MobileNet-V2+0.29-1.0606+1.2980
ShuffleNet-V2 ×0.5-0.25-1.0880+1.5944
ResNet-8+0.75-1.0858-0.5577

CNN and ResNet-8 reduce the full pipeline’s computational cost by 1.1052 and 1.0858 GFLOPs and its latency by 0.8474 and 0.5577 ms, respectively, while maintaining B. MobileNet-V2 and ShuffleNet-V2 also reduce FLOPs but increase measured latency. Inference time depends on memory access, operator overhead, and hardware utilization in addition to FLOPs, particularly for single-sample inference. These results show that the choice of lightweight architecture matters for reducing actual inference latency. We have added this analysis to Appendix S (Table 37). Tables 30 and 31 of the submission are now Tables 35 and 36.

Q4: Source of the accuracy drop at larger client scales

Response: Thank you for suggesting this comparison. The drop comes from the reduced amount of data per client, which affects both the global model and the local Featurizer. The table below adds the original global model (FedAVG, before Eraser) under the same client configurations as Table 16.

Table: FedAVG vs. Eraser at different client scales under BadNet on CIFAR-10 (%)

100 clients200 clients300 clients
ABRAABRAABRA
FedAVG84.630.000.0076.300.000.0071.730.000.00
+Eraser84.9598.9482.0977.3596.9374.0774.1896.6469.40

The global model degrades even without Eraser: from 100 to 300 clients, FedAVG loses 12.90% in accuracy, more than the 10.77% lost with Eraser. Applying Eraser therefore never costs accuracy relative to the global model at any of these client scales. Eraser stays above FedAVG at every scale, and its margin grows from 0.32% to 2.45% as local data shrinks. The defense also holds at every scale, with B above 96.6% and RA within 5% of A. We add these results to Appendix J, where Table 16 of the submission is now Table 17 and includes FedAVG.

Dear Reviewer ABjo,

Thank you for your constructive feedback. We have completed the supplementary experiments addressing your questions, including evaluations of reconstruction accuracy, cross-client feature compatibility, computational overhead, and scalability, and have incorporated the results into the revised manuscript.

Before the discussion period closes, we would appreciate your feedback on whether these results and explanations address your concerns. Please let us know if any points require further clarification.

Thank you again for your constructive review.

Best regards,

The authors

Reviewer_aHXWOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Eraser, a post-training sample-purification defense against backdoor attacks in federated learning. It requires no additional clean or public auxiliary dataset and does not rely on inspecting client updates to identify malicious clients. After the main federated model has been trained, each client trains a client-specific Featurizer and a lightweight Generator. At inference time, the Featurizer maps a potentially triggered input into a latent representation that suppresses trigger-related cues, and the shared Generator reconstructs a sanitized sample for the frozen global model. To mitigate cross client feature space inconsistency under non-IID data, the server samples fixed anchor for each class and normalizes it onto the unit hypersphere. Each client trains its Featurizer to pull local features toward class anchors and push them away from the other anchors.

Critical strengths

1.Eraser avoids identifying malicious clients or modifying the global training process, instead suppressing triggers through semantic reconstruction before inference, which provides strong compatibility.

2.The experimental evaluation is comprehensive, covering multiple datasets, backdoor attacks, and heterogeneous federated settings. The results show that Eraser substantially reduces attack success rates while preserving clean-task accuracy.

Critical weaknesses

1.Since LSTR\mathcal L_{\mathrm{STR}} uses cosine similarity, normalizing a gaussian anchor does not change its similarity to any feature vector, cos(z,a~c)=cos(z,a~c/∥a~c∥)cos(z,\tilde a_c)=cos(z,\tilde a_c/\|\tilde a_c\|). Why, then, does hyperspherical projection improve feature discrepancy over unnormalized random anchors in Figure 4?

2.How do the authors demonstrate that Eraser works by extracting and regenerating the input semantics, rather than simply discarding input details and producing a class-conditioned pattern? This distinction is critical because the reconstructions in Appendix Fig. 6 are largely unrecognizable.

3.Appendix T formulates the actual staged updates as joint SGD. But Assumption 3 only characterizes gradient discrepancies, while Lemma 2 claims bounds on the aggregated parameter drift and the increment of the objective function. Equation (38) cannot directly establish the bounds required for Equation (43), and the first-order terms are not controlled either. Even if these issues were corrected, the result would establish only generic convergence of nonconvex SGD, rather than guarantees of backdoor elimination or utility preservation.

4.To my knowledge, this paper is not the first federated defense strategy based on post-training sample purification [1–3]. The authors should revise their contribution claims accordingly.

[1] FedPurify: Knowledge-Preserving Backdoor Defense with Data-Free Purification in Federated Learning, Xue et al., ACM KDD 2026.

[2] BAERASER: Backdoor Defense with Machine Unlearning, Liu et al., INFOCOM 2022.

[3] FedRecover: Recovering from Poisoning Attacks in Federated Learning Using Historical Information, Cao et al., IEEE S&P 2023.

Questions for authors

see weaknesses

6 comments
AuthorsOpenReview ↗

Response (1/3)

We greatly appreciate your careful review and detailed comments.

W1: Anchor normalization and Fig. 4

Response: Thank you for pointing this out. We apologize for the unclear description in the original manuscript. We compute the score as a dot product after normalizing the features: z^=z/∥z∥2\hat z=z/\|z\|_2 and sc=z^⊤ac/τs_c=\hat z^\top a_c/\tau. Since the anchors are also normalized by Eq. (2), z^⊤ac\hat z^\top a_c equals cosine similarity. We therefore wrote Eq. (4) directly in cosine form without separately showing the feature-normalization step. To make the computation explicit, we have revised Sec. 4.2 to show the feature normalization and write the score as a dot product.

In the “Random Anchors” setting of Fig. 4, feature normalization is retained, but anchor normalization is omitted. The same dot product therefore gives sc=∥a~c∥2cos⁡(z,a~c)/τs_c=\lVert\tilde{a}_c\rVert_2\cos(z,\tilde{a}_c)/\tau. The anchor norm now scales the score, so this setting no longer computes cosine similarity alone. For d=512d=512, the Gaussian anchor norms concentrate around d≈22.6\sqrt{d}\approx22.6, giving effective temperatures τc=τ/∥a~c∥2≈0.0044\tau_c=\tau/\lVert\tilde{a}_c\rVert_2\approx0.0044 when τ=0.1\tau=0.1.

The much lower effective temperature makes the softmax sharper and changes how LSTRL_{\mathrm{STR}} guides feature alignment. Random anchors still give all clients shared class targets, which encourage cross-client alignment, consistent with their lower discrepancy compared with no anchors in Fig. 4. Hyperspherical anchors retain these shared targets while removing the norm factors and giving every class the same angular temperature of 0.1. In Fig. 4, this setting further reduces the discrepancy and achieves the best cross-client consistency among the three settings. We have clarified the anchor settings and their score computations in Appendix D.

W2: Image-specific information in the reconstructions

Response: Eraser is deliberately not designed to restore the input's appearance. The Generator is optimized solely through the classification loss in Eq. (8), without a pixel-level reconstruction loss. Human-recognizable reconstruction is therefore not an explicit training objective, and the appearance of the outputs in Fig. 6 alone does not determine whether they retain input-specific information. We therefore examine this question using representational similarity and instance-retrieval analyses.

We follow the default CIFAR-10 settings and the standard Eraser training procedure. We evaluate the final aggregated shared Generator with all 12 benign clients’ Featurizers on a common, class-balanced held-out set containing 100 images per class. All models remain fixed during evaluation. Scores are averaged across groups within each client and then across clients.

Specifically, we compare the frozen global model’s penultimate-layer features for each original image and its reconstruction. Images are grouped by their true class, or by both their true class and the label predicted by the local classifier. The second grouping controls for differences in discrete predicted labels: a Generator producing a fixed pattern determined solely by the predicted class would yield identical reconstructions within each group, leaving no within-group similarity structure and only chance-level instance retrieval.

Table: Original vs. reconstructed images in the global model's penultimate-layer features

GroupingRSARetrievalCentered retrievalChanceCoverage
Same true class0.59513.83%18.90%1.00%100.00%
Same true and predicted class0.46120.32%30.09%3.62%89.40%

RSA is the Spearman correlation between the pairwise cosine-distance matrices of originals and reconstructions within a group. Retrieval asks whether the original closest to a reconstruction is its own source, and centered retrieval first subtracts each group's mean. Groups with fewer than 10 images are excluded, and coverage is the fraction of images retained.

Within each class, the reconstructions preserve much of the similarity structure among the originals (RSA 0.595), and a reconstruction finds its own source among 100 same-class images 14 to 19 times more often than chance. Both effects remain when the predicted label is also fixed (RSA 0.461, centered retrieval 30.09% against 3.62% chance), so they do not come from the predicted label. Eraser's outputs therefore retain image-specific information beyond the class label, measured in the global model's representation rather than in pixel space. We add this analysis to Appendix M (Table 23).

AuthorsOpenReview ↗

Response (2/3)

W3: Convergence analysis in Appendix T

Response: Thank you for the careful reading of Appendix T, which we have revised on each point. Equation numbers below refer to the revised manuscript. Assumption 3, Eq. (38) of the submission, is now Eq. (45), and the bound of Eq. (43) is replaced by Lemma 2 in Eq. (50) and Theorem 1 in Eq. (57). The analysis no longer treats the staged updates as joint SGD: following Algorithm 1, each client first updates its Featurizer and then, with the Featurizer fixed, its local Generator, and the two parts are analyzed separately because the reconstruction gradient never reaches the Featurizer. Assumption 3 now bounds gradient heterogeneity only at a common Generator, and Lemma 2 derives the drift of the local Generators from their average through the SGD recursions in Eqs. (51)–(56) rather than from Assumption 3 alone. Theorem 1 keeps the first-order term, which Lemma 2 controls, and bounds the effect of the Featurizer updates separately (Assumption 4). For fixed QQ and constants independent of EE, with stepsize η=1/(LEQ)\eta=1/(L\sqrt{EQ}) and E≥9QE\ge9Q, Theorem 2 in Eqs. (64)–(66) bounds the time-averaged expected squared gradient norms at rates O(1/EQ)O(1/\sqrt{EQ}) for each local Featurizer objective and O(1/(EQ)1/4)O(1/(EQ)^{1/4}) for the current-round Generator objectives. Both quantities therefore vanish as EE increases.

Appendix T answers a different question from backdoor removal: whether Eraser’s staged training converges when clients follow Algorithm 1. Backdoor attacks aim to remain stealthy while preserving main-task accuracy, and a backdoored model can still converge normally. Backdoor elimination and utility preservation have already been evaluated experimentally in the manuscript. Malicious clients remain active during Eraser training (Sec. 5.1), and Table 14 shows that Eraser maintains B around 98%–99% and achieves the highest V among the compared methods whether attacks start at 10%, 30%, 70%, or 90% of training.

AuthorsOpenReview ↗

Response (3/3)

W4: Relation to prior post-training defenses

Response: Thank you for the references. After reviewing these works, we found that their defenses focus on model-level purification or recovery. Specifically, FedPurify and BAERASER repair the model, while FedRecover recovers the global model after malicious clients have been identified. Eraser applies purification to the inference samples: it keeps the global model unchanged and purifies each input before passing it to the model.

  • FedPurify is a post-training FL defense that synthesizes surrogate data, inverts triggers, and updates the compromised global model through contrastive feature alignment and self-distillation, so its purification target is the model's parameters.
  • BAERASER recovers trigger patterns and removes the backdoor from the model by gradient-ascent unlearning, and it was designed for centralized training rather than FL.
  • FedRecover requires the malicious clients to be identified beforehand and uses the stored history of global models and client updates to estimate the benign clients' updates and recover the global model.

Eraser needs neither malicious-client identification nor stored training history, and malicious clients may remain active during its training. These works focus on model-level purification or recovery, while Eraser is the first post-training sample-purification defense for FL.

We also compare with the three methods on eight attacks, giving FedRecover the identities of all malicious clients.

Table: Comparison with prior post-training defenses on CIFAR-10 (%)

BadNetDBASignalBlended
ABVABVABVABV
FedAVG86.230.0043.1285.570.0042.7985.510.0042.7686.060.0043.03
+FedPurify85.7912.5149.1586.7150.8268.7786.381.5343.9685.100.0042.55
+BAERASER77.770.0138.8977.850.0138.9372.2423.9248.0877.505.6741.59
+FedRecover86.2492.8289.5385.3299.7092.5186.1599.8893.0285.9699.8292.89
+Eraser87.6898.7693.2287.6798.3793.0287.8999.7793.8387.9599.9093.93
WaNetBppAttack1977Kelvin
ABVABVABVABV
FedAVG84.6622.8953.7884.872.6343.7586.923.2445.0886.812.4644.64
+FedPurify84.3529.4056.8783.313.4943.4082.993.9943.4985.142.6743.90
+BAERASER73.2045.4059.3078.385.4841.9376.7918.7547.7773.157.9740.56
+FedRecover85.1199.3392.2285.0899.9792.5385.1599.2192.1885.9298.4292.17
+Eraser87.8899.5093.6987.9399.4393.6888.0099.2893.6487.8499.1593.50

Eraser achieves the highest V on every attack. FedPurify and BAERASER fail to remove the backdoor (B at most 50.82%), and BAERASER also lowers A by 6.49% to 13.66%. FedRecover comes closest, but its A is 1.44% to 2.85% lower than Eraser's on all eight attacks, even though it is given the identities of all malicious clients. We add this comparison to Appendix N (Table 27) and describe the three methods in Appendix A.

[1] FedPurify: Knowledge-Preserving Backdoor Defense with Data-Free Purification in Federated Learning, Xue et al., ACM KDD 2026. [2] BAERASER: Backdoor Defense with Machine Unlearning, Liu et al., INFOCOM 2022. [3] FedRecover: Recovering from Poisoning Attacks in Federated Learning Using Historical Information, Cao et al., IEEE S&P 2023.

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

This paper points out a gap where backdoor attacks are becoming stealthier and cannot be applied to heterogeneous scenarios. To address the gap, the authors introduce ERASER, which is an inference-time backdoor defense that can remove trigger effects by observing inputs from a latent semantic space and reconstructing sanitized samples before the global model converges into the poisoned space. ERASER introduces a 2-component semantic reconstruction consisting of a Featurizer and a Generator. Experimental results show that ERASER achieves the best defense performance while maintaining clean task accuracy.

Critical strengths
  • This paper points out a gap in backdoor defenses where most of the work relies on the input space with little to no consideration of the semantic latent space. Separating local representation learning from generator aggregation is interesting because it can limit direct poisoning of benign clients. Also, the paper does not assume that malicious participants become honest after the original federated training stage.
  • The experiments are sufficient, with multiple settings and multiple baselines, with high results in every baseline.
Critical weaknesses

Major Weaknesses

  • I have concerns about the convergence proof. In Appendix T, you assume that similar local and global gradients imply that averaging the generators will only slightly increase training loss. However, gradients can be identical while averaging still increases a client’s loss. The proof also does not accurately reflect that the featurizer and generator are trained separately. These issues must be corrected before claiming the stated convergence guarantee.
  • The authors argue that clean and triggered inputs produce similar features, making it difficult for a poisoned generator to distinguish them. However, small differences may still reveal whether a trigger is present, and a malicious generator could potentially learn to exploit them. Although the paper evaluates adaptive attacks, additional analysis is needed to better support the claim that trigger information has been sufficiently suppressed.

Other Concerns

  • I am not sure if the reconstruction stage is necessary since the local featurizer already learns from class labels and has a classifier attached during training.
  • There are some reporting errors found in Tables 3, 7, 9 and 12.
Questions for authors
  • I am wondering whether this defense method introduces any additional considerable latency on top of the federated aggregation time.
  • What is the correct-label accuracy on triggered inputs?
  • Can you describe the hyperspherical feature space claim in more detail and how it is relevant to the current work?
  • Please justify if the reconstruction stage is necessary.
  • Please correct the reporting errors in Tables 3, 7, 9 and 12.
6 comments
AuthorsOpenReview ↗

Response (1/3)

We thank you for the helpful comments and suggestions. Below, we address the concerns with a revised convergence analysis and additional experiments. Unless otherwise specified, all experiments follow the default settings described in the manuscript.

W1: Convergence analysis and separate training stages

Response: Thank you for these comments on the convergence analysis. The revised Appendix T addresses both points. Following Algorithm 1, it models the two stages separately: in each round, a client first updates its Featurizer on LEi\mathcal L_{E_i} and then, with the Featurizer fixed, updates its local Generator, and only the Generators are aggregated. The deviation between the local Generators and their average is no longer inferred from gradient similarity. Lemma 2 derives it from the SGD recursions, using the common Generator at the start of each round, smoothness, bounded gradient noise, and gradient heterogeneity measured at a common Generator (Assumption 3). Theorem 1 then analyzes the global Generator objective along the averaged Generator with the first-order term explicitly controlled, so the bound holds even when aggregation raises an individual client's loss, and it bounds the effect of the Featurizer updates on this objective separately through a bounded-sensitivity assumption (Assumption 4). Theorem 2 then gives the rates. For fixed QQ and constants independent of EE, with stepsize η=1/(LEQ)\eta=1/(L\sqrt{EQ}) and E≥9QE\ge9Q, the time-averaged expected squared gradient norms are bounded at rates O(1/EQ)O(1/\sqrt{EQ}) for each local Featurizer objective and O(1/(EQ)1/4)O(1/(EQ)^{1/4}) for the current-round Generator objectives. Both quantities therefore vanish as EE increases. The slower Generator rate comes from the Featurizer updates, which change the Generator objective between rounds. Once the Featurizers are no longer updated, it returns to O(1/EQ)O(1/\sqrt{EQ}).

W2: Residual trigger information and its exploitation across clients

Response: Thank you for raising this point. Exploiting these differences consistently requires cues that transfer across benign clients’ private Featurizers, since their features are processed by the shared Generator.

We test this transferability directly. The attacker first trains a local Featurizer on clean data for 300 epochs to simulate a benign client. It then freezes the Featurizer and trains linear and MLP probes for 100 epochs each to distinguish features from clean and triggered samples. Each probe is subsequently frozen and applied without further training to features produced by benign clients’ Featurizers, which are trained following the standard Eraser procedure.

Table: Trigger-detection accuracy (%) of the attacker's probes on its own Featurizer and on benign clients' Featurizers under CIFAR-10

AttackLinear: attacker-localLinear: benign-transferMLP: attacker-localMLP: benign-transfer
BadNet93.2246.0599.4150.22
DBA54.7751.9256.6950.19
Signal97.2750.0098.2346.65
Blended98.6050.0098.8949.88
WaNet50.0050.0050.0050.00
BppAttack94.4050.0696.6151.51
197792.4850.0096.0250.00
Kelvin94.6250.0096.7649.57

On the attacker's own features, both probes exceed 92% on six of the eight attacks (up to 99.41%), so small differences can indeed reveal the trigger. These cues are specific to that Featurizer: on benign clients' features, the same probes score 46.05% to 51.92%, the chance level on the balanced evaluation set (entries of exactly 50.00% are probes that assign every sample to one class). The anchors and LSTRL_{STR} supervise each sample only through its class and place no constraint on the trigger-induced shift, which therefore differs across independently trained Featurizers. This is what Sec. 4.2 means by the Generator lacking reliable trigger-specific cues: the cues exist, but not in a form a shared Generator can rely on across benign clients, whose Featurizers stay private and are never aggregated.

We add the probe analysis to Appendix O (Table 30) of the revised manuscript.

AuthorsOpenReview ↗

Response (2/3)

W3 & Q4: Necessity of the reconstruction stage

Response: The reconstruction stage is necessary because the auxiliary head HiH_i is trained only on the client's local data, whereas the Generator lets the client's features be classified by the frozen global model, which is trained on all clients' data. Its goal is to convey task-relevant information to the global model, not to recover a photorealistic image. The table compares the three prediction paths. RA is the accuracy on triggered inputs with respect to their true labels.

Table: Prediction paths on CIFAR-10, averaged over the eight attacks in Q2 (mean ±\pm std over three seeds)

SettingPredictionA (%)B (%)RA (%)Semantic retention (%)Induced target flip (%)
Local classificationHi(Fi(x))H_i(F_i(x))71.01 ±\pm 2.7298.82 ±\pm 1.0952.59 ±\pm 0.2268.48 ±\pm 2.390.58 ±\pm 0.16
FedAVGG(x)G(x)86.53 ±\pm 2.383.87 ±\pm 2.533.65 ±\pm 1.474.29 ±\pm 1.7695.61 ±\pm 0.66
Full EraserG(R(Fi(x)))G(R(F_i(x)))88.30 ±\pm 1.8599.07 ±\pm 0.7268.15 ±\pm 1.3475.49 ±\pm 3.240.53 ±\pm 0.28

Semantic retention and induced target flip are measured on non-target samples classified correctly when clean, as the fractions that stay correct or switch to the target class after triggering. Predicting with HiH_i alone suppresses the backdoor but limits A to 71.01%. Passing the same features through the Generator and the global model raises A by 17.29% and RA by 15.56% with comparable B, and keeps more predictions correct under triggering (75.49% vs. 68.48%) at an equally low target flip. Reconstruction is therefore what lets Eraser retain the global model's accuracy without inheriting its backdoor. We add this comparison to Appendix M (Table 24).

W4 & Q5: Reporting errors in Tables 3, 7, 9, and 12

Response: Thank you for pointing these out. The errors came from computing V=(A+B)/2\mathcal V=(\mathcal A+\mathcal B)/2 in Eq. (11) by hand, and we have corrected the affected entries. The underlying A and B values are unchanged, and the corrections do not change any conclusion or the leading performance of Eraser and its combinations in these tables. Since the revised appendix adds new tables, Table 12 of the submission is now Table 13.

Q1: Additional training and inference latency

Response: On CIFAR-10, Eraser adds almost no aggregation time: its server-side aggregation takes 0.33 s, compared with 10.95 s for the main federated training. Its main cost is a client-side post-training stage.

Table: Training and aggregation time on CIFAR-10 (s)

StageClient-side optimization time (s)Server-side aggregation time (s)Total runtime (s)
Main federated training4225.3710.954236.32
Eraser post-training1185.850.331186.18
Complete pipeline5411.2211.285422.50

The complete pipeline takes 1.28× the runtime of the main training (i.e., 28.0% additional runtime), almost all of it due to client optimization. We add this breakdown to Appendix S (Table 39). At inference, the Featurizer and Generator raise latency from 3.16 ms to 5.75 ms per sample (+82.2%), as reported in Table 36.

Table: Per-sample inference latency under the default ResNet-18 setting

MetricGlobal model onlyEraser + global modelAdditional overhead
Inference latency (ms/sample)3.15775.7539+2.5962 ms (+82.2%)
AuthorsOpenReview ↗

Response (3/3)

Q2: Correct-label accuracy on triggered inputs

Response: Thank you for this suggestion. Eraser does more than divert triggered samples from the target class: it largely restores their true labels. We report recovery accuracy (RA), i.e., the accuracy of triggered samples on their ground-truth labels. "Ideal" denotes a FedAVG global model trained without malicious clients and evaluated on the same triggered inputs, so its RA shows how much the trigger itself damages the input.

Table: RA (%) on CIFAR-10

AttackIdealFedAVG+Eraser
BadNet78.280.0086.24
DBA82.170.0086.02
Signal42.440.0050.92
Blended33.030.0046.16
WaNet85.1121.0386.85
BppAttack50.212.5857.40
197767.671.7872.11
Kelvin49.852.3656.94

FedAVG's RA is near zero on seven of the eight attacks, whereas Eraser's RA exceeds even the Ideal model's on all eight, by 1.74% to 13.13%. RA is lower on Signal, Blended, BppAttack, and Kelvin because these are global triggers that alter the content of the entire image, which lowers the Ideal model's RA as well. As shown in the table below, when the blending ratio χ\chi of the Blended trigger in Eq. (20) increases from 0.4 to 0.9, the Ideal model’s RA falls from 21.46% to 13.69%, while Eraser remains 10.80–19.04% higher across all six settings.

Table: RA (%) under Blended triggers with blending ratio χ\chi

0.40.50.60.70.80.9
Ideal21.4620.7718.7615.5213.6713.69
FedAVG0.000.000.000.000.000.00
+Eraser40.5034.9534.3728.8229.3824.49

We define RA in Appendix B and report these results in Appendix G (Table 12).

Q3: Meaning and relevance of the hyperspherical feature space

Response: Thank you for this question. "Hyperspherical" refers to the class anchors rather than to the features. For each class cc, we sample a fixed Gaussian vector a~c∼N(0,Id)\tilde a_c\sim\mathcal N(0,I_d) and use its direction ac=a~c/∥a~c∥2∈Sd−1a_c=\tilde a_c/\Vert\tilde a_c\Vert_2\in\mathbb S^{d-1} as the class anchor in Eqs. (1)–(2), and the structural loss in Eq. (4) compares each feature's direction with these anchors through cosine similarity. The features passed to the Generator are not required to have unit norm. These shared directions matter because each client trains its Featurizer locally while all clients share one Generator. Since the loss is applied to every sample, it pulls the features of each class toward the same direction on every client.

We conduct this analysis with 20 clients on CIFAR-10, using the same data partitions across the two conditions. Each Featurizer undergoes 300 epochs of clean local training and is then frozen, followed by 50 communication rounds of shared Generator training. The two Generators start from identical weights and use the same frozen global model. Evaluation uses a common, class-balanced held-out set containing 100 images per class. Our cross-client analysis confirms this: with anchors, the class-conditional MMD2\mathrm{MMD}^2 between clients falls by 40.53%, and still by 33.60% after removing class-mean offsets, so the anchors align within-class distributions and not only class means, accompanied by a 3.02% gain in the joint correct rate, i.e., the fraction of images that two clients both classify correctly when each passes the image through its own Featurizer and the shared Generator.

Table: Effect of shared semantic anchors on cross-client feature compatibility under BadNet on CIFAR-10 (relative change for feature distances and absolute change in % for joint correct rate)

MetricRaw MMD2\mathrm{MMD}^2Centered MMD2\mathrm{MMD}^2Centered MMD2\mathrm{MMD}^2 (classes seen by both clients)Δ\Delta joint correct rate
Change with anchors−40.53%−33.60%−38.75%+3.02

We clarify the meaning of "hyperspherical" in Appendix P and add the cross-client analysis to Appendix D.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers and Summary of Updates

Dear Area Chair and Reviewers,

Before summarizing our revisions, we would like to highlight Eraser’s central contribution: the first post-training sample purification in FL while keeping the global model unchanged, without requiring malicious-client identification. This is unique from the previous methods.

We addressed the reviewers’ concerns through additional experiments, clarifications, and a revised convergence analysis.

  • Reviewer JDKs acknowledged our clarifications and maintained their positive rating.
  • We are concerned about how Reviewer aHXW handled the final assessment. With approximately two hours left in the discussion, they lowered their rating and raised the absence of true-label accuracy as a concern. Appendix G, Table 12 of the revised manuscript already reported that metric. We responded to each point. In our view, reducing the rating while identifying an existing result as missing was unfair to the submission. We ask the Area Chair to review our follow-up before weighing the revised rating.
  • We have not received further feedback from Reviewer ABjo. We respectfully ask the Area Chair to consider the revised manuscript and these follow-up responses alongside the original reviews. We summarize the main updates below.

Best regards,

Authors of Submission 12749

- Revised convergence analysis. Appendix T now models the separate training stages, derives Generator drift from SGD recursions, and controls the first-order terms and Featurizer-induced objective changes. We distinguish optimization convergence from guarantees of backdoor removal or utility preservation.

- Correct-label recovery. We added recovery accuracy (RA) to measure correct-label prediction on triggered inputs. Eraser achieves 46.16% to 86.85% RA across eight attacks and 86.26% under ISSBA, compared with 0.13% for FedAVG under ISSBA (Appendix G).

- Necessity of reconstruction. A three-seed ablation shows that the complete pipeline improves clean accuracy by 17.29 percentage points and RA by 15.56 percentage points over local classification, while maintaining comparable backdoor resistance (Appendix M).

- Input-specific information. We clarified that pixel-level reconstruction is not the training objective. With true and predicted classes fixed, RSA is 0.461 and centered retrieval is 30.09% versus 3.62% chance, supporting input-specific information beyond fixed class patterns in the global model’s feature space (Appendix M).

- Anchor normalization and feature alignment. We clarified how unnormalized anchor norms change the effective temperature. Under BadNet, shared anchors reduce centered cross-client feature discrepancy by 33.60% and improve the pairwise joint correct rate by 3.02 percentage points (Appendix D).

- Transferability of trigger cues. We added linear and MLP trigger-detection probes. When transferred without retraining to benign clients’ private Featurizers, their accuracy is 46.05% to 51.92% on a balanced test set, showing limited direct transfer of the tested detection rules (Appendix O).

- Comparisons with related defenses. We added comparisons with FedPurify, BAERASER, and FedRecover across eight attacks, giving FedRecover the identities of all malicious clients. Eraser achieves the highest composite score V in every evaluated attack without requiring these identities (Appendix N).

- Training and inference overhead. We report approximately 28% additional total training runtime, including 0.33 seconds of server aggregation. Both training and inference costs are explicitly reported (Appendix S).

- Lightweight models and scalability. We evaluated five lightweight Featurizers and matched FedAVG baselines for 100, 200, and 300 clients. CNN and ResNet-8 reduce inference latency. Eraser maintains higher clean accuracy than FedAVG at every evaluated client scale (Appendices J and S).

We thank the reviewers and Area Chair for their time and consideration. We will incorporate all supporting justifications into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.