acceptodds
Withdrawn from ICLR 2027

Residual-Direction Filtering for Backdoor Defense in Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Malicious clients in federated learning can implant backdoors through poisoned updates, and server-side defenses must filter such updates without clean data or knowledge of the attackers. Detection-based defenses increasingly score clients on parameter subsets chosen by dedicated mechanisms, such as importance ranking or critical-layer selection, yet it is unknown when scoring a subset yields the same filtering decisions as scoring all parameters. We propose CoPR, which scores each client on a fixed parameter subset by the direction of its cross-client mean-centered residual relative to the coordinate-wise median residual, filters clients with a median-absolute-deviation threshold, and aggregates the clipped full updates of retained clients. We give sufficient conditions under which residual directions separate malicious from benign clients, show that a parameter subset preserves every filtering decision whose margin exceeds the subset-induced score perturbation, and bound the single-round aggregation error under missed detections, false rejections, and clipping. Consistent with this criterion, subsets holding only 0.78% of the parameters match full observation on federated diffusion models at every tested location with 13–54× lower detection time, whereas non-IID classification requires full observation. Across five attacks on CIFAR-10 diffusion models, CoPR attains the lowest average FID while suppressing private-image reconstruction as well as AlignIns and remains effective against a subset-aware variant of AdaSCP; in non-IID classification, CoPR-full reduces the average attack success rate to 6.86% at 82.78% clean accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Backdoor Defense, Compact Parameter Residuals, Robust Aggregation

Reviews

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

This paper proposes the CoPR algorithm, a server-side backdoor defense method for federated learning. The defender server performs cross-client mean centralization on the parameters selected by each client, calculates the cosine deviation fraction of the residual direction relative to the median residual direction for each coordinate, then filters out suspicious clients based on a threshold for the absolute deviation from the median, and finally trims and weight-aggregates the complete updates from the remaining clients according to the median norm. Furthermore, the paper analyzes the sufficient conditions under which the residual direction can distinguish malicious clients, proves that a subset of parameters can preserve all filtering decisions where the confidence margin exceeds the fractional perturbation caused by that subset, and provides upper bounds on single-round aggregation error under false acceptance, false rejection, and clipping scenarios.

Critical strengths
  1. This paper addresses a meaningful and specific problem: When the defender’s server lacks clean reference data and does not know the number of malicious clients, is it possible to obtain the same filtering decision as would be achieved by observing all model parameters by observing only a small number of them, thereby filtering out backdoor updates at a lower detection cost?
  2. The motivation regarding residual directions is compelling: when malicious clients are few in number, their updates share a common offset, and benign residual directions are relatively concentrated, mean centralization may cause the residuals of the two groups to point in opposite directions. The paper not only provides this explanation but also explicitly states the conditions required for directional separation and subset decision stability.
  3. The evaluation not only covered various poisoning strategies but also simultaneously examined attack suppression and the quality of normal tasks. The study achieved good results for the five attacks on the tested CIFAR-10 diffusion models and the four attacks in the IID classification experiments, and truthfully reported the failure of the default small-parameter slicing in non-IID classification.
Critical weaknesses
  1. The experimental results in the paper do not yet fully validate the effectiveness of the proposed residual direction mechanism. Table 6 in the ablation experiments only examines ablation for filtering and clipping; it does not separately test the necessity of mean centralization, coordinate-wise median reference directions, and cosine scores under the same thresholds and aggregation processes. Consequently, it is impossible to determine whether the performance stems from the newly proposed mechanism or from the general filtering process and experimental configuration.
  2. Although the authors claim that CoPR does not require prior knowledge of the number of malicious clients, its detection theory explicitly relies on conditions such as the proportion of malicious clients being less than 1/2 and the directional concentration of normal clients. Furthermore, the fact that the defender does not require knowledge of the reference set or triggers is consistent with the assumptions of most current backdoor defense methods (e.g., [1], [2], [3], [4], [5]); therefore, this cannot be considered an innovation of this paper. [1]Nguyen et al., “FLAME: Taming Backdoors in Federated Learning,” USENIX Security, 2022. [2]Rieger et al., “DeepSight: Mitigating Backdoor Attacks in Federated Learning Through Deep Model Inspection,” NDSS, 2022. [3]Jebreel and Domingo-Ferrer, “FL-Defender: Combating Targeted Attacks in Federated Learning,” Knowledge-Based Systems, 2023. [4]Huang et al., “Multi-Metrics Adaptively Identifies Backdoors in Federated Learning,” ICCV, 2023. [5]Ma et al., “A Stability-Enhanced Dynamic Backdoor Defense in Federated Learning for IIoT,” IEEE Transactions on Industrial Informatics, 2024.
  3. This paper adopts a hard filtering strategy based on residual direction statistics and a MAD threshold to directly exclude clients whose cosine deviation scores exceed the threshold. This strategy may be relatively sensitive to heterogeneity in client data. However, the non-IID scenarios tested in this paper are relatively simple, primarily simulating Dirichlet partitions, i.e., label skewness. More realistic non-IID scenarios—such as feature distribution shifts, imbalanced client sample sizes, temporal heterogeneity, conceptual shifts, or partial client participation per round—were not tested.
  4. The non-IID classification comparison only evaluated the default 5,120-parameter slice against the full version with approximately 6.57 million parameters; intermediate scales were not tested. This does not sufficiently prove that non-IID classification must observe all parameters.
  5. The scope of the attacks does not fully cover the dataset. The main table of the five diffusion attacks is based solely on CIFAR-10, and the cross-dataset results for CelebA primarily focus on the AdaSCP diffusion attack; therefore, they cannot be generalized to other classification datasets or models.
Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer GRuY

We thank the reviewer for the detailed comments. Our responses follow.

W1. Table 6 shows that the filtering stage is needed: adding it to clipping alone improves both FID and Stolen-MSE. We agree that it compares filtering against no filtering and does not isolate the score. On the coordinate-wise median: after mean-centering, the residuals sum to zero, so a mean reference is identically zero and its cosine is undefined; the median is the natural robust nonzero reference. To test centering and the cosine, we kept the MAD threshold (λ = 2), clipping, aggregation, and the observed subset fixed, and replaced only the score. On CIFAR-10 diffusion under AdaSCP (three training seeds), the attacker's margin is large (Fig. 6): all four scores reject it in every round with comparable FID and Stolen-MSE, and the norm-only score has a higher benign false-rejection rate (5.08 ± 0.44%) than the other three (3.58–3.89%). The differences are larger under non-IID classification (Table 3b protocol, seed 0, one fixed subset of 2% of the weights; ASR in %):

ScoreBadNetDBANeurotoxinPGD
Centered cosine (CoPR)1.531.391.431.32
Centered L263.268.931.3260.34
Norm only59.3621.6128.8159.08
Uncentered cosine70.081.611.2163.10

Clean accuracy is 82.5–84.8% in all cells. Only the centered cosine suppresses all four attacks.

W2. (a) These are two different things. CoPR does not take the number of attackers as an input: the retained set is determined by the score distribution of each round through the MAD rule, whereas Krum and Multi-Krum require this number. A benign majority and the angular concentration of benign residuals are assumptions of Theorem 1, which gives a sufficient condition under which detection is guaranteed; the method runs without them. A benign majority is also assumed by defenses that do not require the attacker count, e.g., FLAME (Nguyen et al., 2022).

(b) We do not present the absence of clean reference data, trigger knowledge, or the attacker count as a novelty. The sentence in our contribution list states the conditions under which CoPR operates, which it shares with the defenses the reviewer lists. Our contributions are the residual-direction score and the characterization of when a fixed parameter subset preserves the filtering decisions of full observation (Theorem 2).

W3. We agree that hard filtering is sensitive to heterogeneity; Table 3b and Fig. 5 show this. On partial participation: CoPR decides from the uploads of each round and keeps no client-specific state, so it applies when clients are sampled. With 100 clients, 20 sampled per round, and 20% attackers (IID partition, BadNet, 200 rounds, three seeds), CoPR-full reduces ASR from 62.21 ± 11.44 to 2.81 ± 0.88. Beyond this, our evaluation of heterogeneity is limited to Dirichlet label skew (β from 0.1 to 1.0 in Fig. 5), as in the closest prior defenses; feature shift, sample-size imbalance, temporal heterogeneity, and concept drift are outside the scope of this paper.

W4. We agree. We ran intermediate budgets under the Table 3b protocol (seed 0), with fixed subsets drawn at random from the convolutional weights (ASR in %):

ObservationBadNetDBANeurotoxinPGDAvg. ASRACC
FedAvg55.5726.1736.7944.7040.8184.35
CoPR-full1.441.501.671.321.4882.58
Conv 2%1.381.641.601.411.5183.41
Conv 5%1.281.361.371.221.3183.29

Observing 2% or 5% of the convolutional weights matches CoPR-full on all four attacks.

W5. Our evaluation covers two task families, federated diffusion generation and federated classification. Within each task, the scope is as the reviewer describes: the five-attack comparison follows the DataStealing protocol on CIFAR-10; on CelebA we evaluate only AdaSCP because federated diffusion training is costly; and classification uses CIFAR-10 with ResNet-9.

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

This paper introduces a backdoor defense for federated learning (FL). It is a detection-based defense that first locates a small subset of parameters uploaded by clients, computes the cross-client parameter residuals, and uses thresholding with median-absolution-deviation for filtering out malicious uploads, followed by clipped aggregation to suppress the remaining ones with a large magnitude that might dominate the learning. The effectiveness of the proposed method, CoPR, has been validated both theoretically and empirically on two tasks: image generation and image classification.

Critical strengths
  • The intention to speed up the detection process at the server-side is important.
  • Theoretical analysis is provided to support the design.
  • Multiple tasks (generation and classification) are considered.
Critical weaknesses
  • The methodology requires all clients to participate in each round, which is uncommon in FL. Also, the experiments seem to always have malicious participants to participate in each round. It is unclear what will happen if the FL system has no attacker at all or the random client selection selects all benign clients to participate. CoRP may remove benign clients in that case.
  • The experiment setting is weak, and the results are not impressive. First, the experiments were conducted at a small scale, with only 5 clients for generation and 20 clients for classification on a toy dataset (CIFAR-10). Second, as reported in Table 1, the improvement made by CoRP is inconsistent. The authors should consider evaluating CoRP on more realistic datasets and setup, following the standard in FL research.
  • The presentation could be further improved.
Questions for authors

Please refer to the weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer hjha

W1: Client participation and attack-free rounds

We thank the reviewer for these two points.

(a) CoPR does not require full participation. The filtering decision in each round depends only on the uploads received in that round: centering, the coordinate-wise median reference, the threshold, and the clipping radius are all computed over the current participants, and CoPR keeps no client-specific state across rounds. The misunderstanding comes from our problem setup, which defines C\mathcal{C} as all clients and says the server broadcasts "to all clients". C\mathcal{C} should be read as the set of clients participating in the current round. Full participation in our experiments follows the protocols of prior work (see W2); it is not a requirement of the method.

We also ran a cross-device experiment: CIFAR-10, IID partition, 100 clients with 20 sampled per round, 20% malicious clients, BadNet, 200 rounds, three seeds.

MethodACC (%)ASR (%)
FedAvg78.16 ± 0.8362.21 ± 11.44
CoPR-full78.72 ± 0.252.81 ± 0.88

CoPR-full remains effective under client sampling.

(b) Attack-free training. Like other filtering-based defenses (for example, Multi-Krum discards a fixed fraction of clients in every round whether or not an attack is present), CoPR can reject benign clients when no attacker is present. We measured the cost:

  • Diffusion (CIFAR-10, 5 clients, no attacker, single run): the benign false-rejection rate is 11.73%. The final model is unaffected: FID is 6.19 vs. 6.23 for FedAvg.
  • Classification (the cross-device setting above, no attacker, three seeds): accuracy is 79.57 ± 0.33 (FedAvg) and 79.82 ± 0.28 (CoPR-full).

W2: Experimental scale

Our experimental scale follows the protocols of prior work. The diffusion experiments use the setting of DataStealing (Gan et al., 2024), with five clients and one attacker. The classification experiments use the setting common to AlignIns (Xu et al., 2025) and related work, with 20 clients and four attackers. Beyond CIFAR-10, the diffusion experiments also include CelebA at 64×64. In addition, we ran a cross-device experiment with 100 clients and 20 sampled per round; the results are in our response to W1.

On Table 1: CoPR's results are nearly constant across the five attacks (FID 6.17–6.22, Stolen-MSE 14.24–14.25 × 10⁻²), and it attains the lowest average FID (6.18).

Reviewer_epz7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CoPR, a server-side backdoor defense for Federated Learning that operates without clean validation data or a known corruption budget. In each round, CoPR extracts a fixed parameter subset, mean-centers across clients, scores each client by its cosine distance to the coordinate-wise median residual, filters out clients above an adaptive MAD-based threshold, and clips full model updates to the median norm of retained clients before weighted aggregation. The authors prove single-round guarantees for conditional detection (Theorem 1), subset decision stability (Theorem 2), and aggregation error (Theorem 3). Experimental results are validated on federated diffusion (CIFAR-10, CelebA) and ResNet-9 classification (CIFAR-10).

Critical strengths
  1. This paper shows a very interesting observation where scoring a small parameter subset matches full-model filtering whenever the gap between benign and malicious updates exceeds the sampling noise. This explains why tiny subsets succeed on diffusion models (with a 13-54x speedup) but fail under classification label skew.
  2. Without needing clean server data or knowing the number of attackers, the method outperforms recent baselines like AlignIns and Lockdown on both CIFAR-10 diffusion and IID classification. I appreciate the inclusion of negative results in the manuscript.
Critical weaknesses
  1. In this paper, the theory claims that inspecting a small parameter subset can produce the same filtering decisions as inspecting the full model. But I do not see any experimental results supporting this claim. Also, it is unclear to me how practical it is to choose a sufficient number of parameters.
  2. The current attack scenario is insufficient. Assuming the attacker knows which parameters CoRP inspects is okay, but the attacker's modification often makes it much less effective against standard FedAvg. I expect to see a stronger attacker test that can show the robustness of the proposed defense.
Questions for authors
  1. On non-IID ResNet9 (Table 3b), does sampling 1-5% of the parameters randomly or from convolutional layers reduce the attack success rate without requiring full-model inspection?
  2. Does the defense hold under multi-attacker collusion on diffusion (e.g., two attackers among ten clients), and does the drop at the 0.2 heterogeneity level in Fig. 5 persist across three seeds?
  3. What does the symbol beside Multi-rum's CelebA FID values (5.90 and 6.33) in Table 2 mean?
  4. I found an inconsistency in the Stolen-MSE changes from 16.40 (Table 2) to 1.14 (Table 11). Can you justify these changes?
  5. As mentioned in Appendix A.1, 3 run experiments will report the mean and std. dev. Please update Table 2 to have a consistent results representation.
1 comment
AuthorsOpenReview ↗

Response to Reviewer epz7

We thank the reviewer for the careful reading. Our responses follow.

W1. The comparison is in Table 4 and Table 13; we did not link these tables to Theorem 2, which made it easy to miss. On CIFAR-10 diffusion under AdaSCP, Table 4 runs CoPR with full observation and with seven subsets of 0.78% of the coordinates (four layer locations, three random draws), each in a separate run. Every subset rejects the attacker in all 300 rounds, as full observation does, with benign false-positive rates of 2.25–3.83% against 3.50%. Table 13 repeats the default subset against full observation over three training seeds: FNR is 0.00 ± 0.00 for both, and FPR is 3.58 ± 0.36 vs. 3.22 ± 0.25.

W2. The attack in our main evaluation is itself adaptive: AdaSCP (Gan et al., 2024) concentrates the poisoned update on a few critical parameters and rescales it to evade distance-based aggregation, and CoPR defends against it (Table 1: Stolen-MSE 14.24 × 10⁻², vs. 1.17 × 10⁻² without defense). As a further attacker test, we also evaluated CoPR against Batman (He et al., CVPR 2026), a recent backdoor attack that aligns malicious updates with benign knowledge to evade detection without weakening the attack. On CIFAR-10 classification (ResNet-9, 20 clients, 4 attackers, three seeds), ASR under FedAvg is 76.27 ± 9.74 (IID) and 82.78 ± 5.64 (non-IID); CoPR-full reduces it to 2.57 ± 0.06 and 1.85 ± 0.36.

Q1. We ran the Table 3b protocol (seed 0) with fixed subsets drawn at random from the convolutional weights (ASR in %):

ObservationBadNetDBANeurotoxinPGDAvg. ASRACC
FedAvg55.5726.1736.7944.7040.8184.35
CoPR-full1.441.501.671.321.4882.58
Conv 2%1.381.641.601.411.5183.41
Conv 5%1.281.361.371.221.3183.29

Observing 2% or 5% of the convolutional weights matches CoPR-full on all four attacks.

Q2. We ran CIFAR-10 diffusion with two AdaSCP attackers among five clients (non-IID, 300 rounds, seed 42):

DefenseFID ↓Stolen-MSE (×10⁻²) ↑TPR (%)FPR (%)
FedAvg95.42collapsed——
Multi-Krum12.729.80——
CoPR6.8514.09100.00.0

Without defense, training collapses. CoPR rejects both attackers in every round with no benign rejection. Our classification experiments also use multiple attackers (4 of 20 clients in Table 3).

Q3. The symbol ‡ indicates that the lowest FID in the column should be read together with its Stolen-MSE: in these two columns Multi-Krum has the lowest FID, but its Stolen-MSE (16.50 and 15.60) is below that of CoPR (19.27 and 18.72), i.e., it suppresses private-image reconstruction less. The caption omitted this definition; we have added it in the revised version.

Q4. The Table 2 entry was an error: the value (FID 5.78, Stolen-MSE 16.40) was entered from a different run by mistake. Both entries report non-IID CelebA under AdaSCP without defense, and the correct values are FID 9.05 and Stolen-MSE 1.14, as in Table 11; the low Stolen-MSE shows that the attack succeeds without defense. We have corrected Table 2.

Q5. Appendix A.1 reports mean ± SD only for experiments with three runs. Table 2 does not have three runs for every entry, so it reports no SD.

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

Related papers

Discussion (0)

Sign in to comment.