acceptodds
Under review as a conference paper at ICLR 2027

TIMBRE: Backdoor Defense in Federated Learning via Layer-wise Spectral Energy Distributions

Abstract

Federated learning is vulnerable to backdoor attacks, and strong data heterogeneity lets malicious updates hide among benign updates that already differ. Advanced backdoors mimic benign updates in norm, distance, or overall direction, weakening defenses built on these statistics. We observe that backdoored updates can differ in how their energy is allocated within a layer even at comparable magnitudes. We therefore propose TIMBRE, a server-side defense that uses only individual client updates, without labels or trusted reference data. TIMBRE computes each update's normalized spectral energy distribution layer by layer, which is invariant to overall scale. TIMBRE compares an update's distribution with those of the other clients in the same round, and the further it departs beyond their typical spread, the stronger the backdoor evidence. Rather than filtering or removing clients, TIMBRE grades its restrictions by the strength and persistence of the evidence: it suppresses modifications lacking support from other clients, keeps useful contributions, and lifts restrictions as evidence fades. In experiments on multiple datasets and models under five representative attacks, TIMBRE keeps every attack success rate in the main setting below 1.4% with a clean-accuracy cost of at most 1.5 percentage points; it outperforms the eight evaluated defenses overall and keeps the attack success rate low as data heterogeneity and the malicious participation ratio vary.

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

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JBmoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TIMBRE, a server-side defense against backdoor attacks in federated learning based on layer-wise normalized spectral energy distributions. The key observation is that malicious and benign updates may exhibit different intra-layer energy allocation patterns, even when their overall magnitudes are similar. TIMBRE uses Hellinger distance and robust statistics to compute spectral anomaly scores, accumulates evidence across communication rounds to identify persistent anomalies, and combines coordinate-level filtering with dynamic aggregation weight adjustment to suppress suspicious updates.

Critical strengths
  1. Rather than relying on update norms, distances, or overall directions, TIMBRE characterizes the internal structure of model updates through layer-wise normalized spectral energy distributions, providing a different perspective for detecting malicious updates that may be difficult to distinguish using conventional geometric statistics.

  2. TIMBRE operates entirely on the server side, without requiring access to clients' local data or modifications to their local training procedures.

  3. The paper is clearly structured and easy to follow.

Critical weaknesses
  1. The paper argues that backdoor training tends to concentrate update energy in a few singular components. However, benign clients with highly skewed class distributions may exhibit similar behavior. The paper therefore does not establish whether this feature remains reliably discriminative under Non-IID conditions.

  2. MARS (NeurIPS 2025) has already explored singular-value-related backdoor energy for malicious client detection in federated learning. Although TIMBRE uses the full normalized spectral energy distribution, the paper does not demonstrate its additional advantages over simpler spectral statistics, such as the largest singular value, spectral entropy, or effective rank.

  3. TIMBRE assigns a strongly restricted aggregation weight coefficient of 0.01 to suspicious clients, which may itself substantially suppress backdoor updates. Although Table D7 provides a component ablation study, it does not compare different anomaly scoring methods under the same suppression mechanism. It is therefore unclear whether the performance gains primarily come from spectral energy-based detection or the subsequent aggressive update suppression.

  4. Section 2.2 claims that deployment requires neither trusted reference data nor an attack-free training phase. However, Appendix C.4 uses a separate 50-round attack-free experiment to determine the thresholds, with different thresholds adopted for different datasets and models. Without experiments on threshold transferability across data distributions, the practical applicability of TIMBRE remains unclear.

  5. PSM-BadNets only attempts to evade detection by mimicking benign spectral energy distributions. It does not jointly optimize against TIMBRE's historical anomaly accumulation and coordinate-level same-sign support mechanisms.

  6. The proposed method involves several hyperparameters, including spectral scoring parameters, anomaly detection thresholds, coordinate filtering proportions, and aggregation weight coefficients. Although Appendix C.4 provides parameter settings and threshold calibration rules, a systematic sensitivity analysis is lacking.

  7. Appendix C.1 indicates that the main experiments continue training from a global model that has already undergone many training rounds, rather than starting from random initialization. Since the spectral structure of updates may differ substantially between early and late training stages, the current results do not establish TIMBRE's effectiveness throughout the complete federated learning process.

Questions for authors
  1. Could the authors report the spectral anomaly score distributions of benign and malicious clients, together with AUROC and false positive rates, under different levels of data heterogeneity?

  2. MARS has already used singular-value-related backdoor energy for malicious client detection. What additional advantages does TIMBRE's full normalized spectral energy distribution offer over simpler spectral statistics, such as the largest singular value, spectral entropy, or effective rank? Could the authors provide comparisons within a unified detection framework?

  3. Could the authors provide controlled experiments to demonstrate that the performance improvements are indeed attributable to the spectral energy features, rather than the subsequent update suppression mechanism?

  4. Appendix C.4 uses a separate 50-round attack-free experiment to determine thresholds, with different thresholds for different datasets and models. Does TIMBRE require dataset- and model-specific attack-free calibration?

  5. The authors should consider stronger adaptive attacks that jointly optimize spectral distribution similarity and coordinate-level sign consistency, or distribute malicious updates across communication rounds to evade historical anomaly accumulation.

  6. The authors should provide sensitivity analyses for key hyperparameters (e.g., rr, κ\kappa, ρcoord\rho_{\mathrm{coord}}, qq, msm_s, and mcm_c) to evaluate whether the method maintains stable performance across different parameter settings.

  7. The authors should include complete federated learning experiments starting from random initialization and report ASR, ACC, and spectral anomaly detection performance at different training stages.

13 comments
AuthorsOpenReview ↗

Response to Reviewer JBmo (1/8)

We thank the reviewer for the detailed comments and have added all the suggested experiments. An overview by question comes first:

  • W1 (discriminability under heterogeneity): we report the evidence distributions, AUROC, and FPR at three heterogeneity levels and group benign clients by class skew. The AUROC is at least 0.921, and skewed benign clients are not more likely to be flagged.
  • W2 (simpler spectral statistics): we replace the score in the same pipeline with the largest singular-value energy, which fails in most settings, or with spectral entropy or effective rank, which degrade markedly under stronger heterogeneity or more malicious clients; the worst final ASR of TIMBRE over eight settings is at most 2.23%.
  • W3 (evidence vs. down-weighting): keeping the response mechanism and replacing only the evidence with random numbers leaves the final ASR at 73%–101% of that of FedAvg without defense in the same setting.
  • W4 (per-dataset calibration): the thresholds are determined once per dataset–model combination before deployment; reusing the CIFAR-10 thresholds on CIFAR-100 changes the ASR by less than 0.5 percentage points.
  • W5 (stronger adaptive attacks): ten configurations that, on top of spectral mimicry, also evade same-sign support and cross-round accumulation; the final ASR of TIMBRE is 0.46%–5.80%.
  • W6 (hyperparameter sensitivity): 37 values of the six parameters you named, varied one at a time; the final ASR is 0.20%–2.47%, and mcm_c matters more when the evidence is weaker.
  • W7 (training from random initialization): runs from random initialization and from checkpoints at rounds 50, 200, and 600, reporting the ASR, ACC, and AUROC at each stage; the final ASR is 0.89%–4.96%, with an accuracy cost from random initialization.

W1 / Q1: Benign clients with highly skewed class distributions may also concentrate update energy in a few singular components, so it is unclear whether this feature remains reliably discriminative under non-IID conditions. Could the authors report the spectral anomaly score distributions of benign and malicious clients, with AUROC and false positive rates, under different levels of data heterogeneity?

The new Appendix D.4 reports the separation achieved by the current-round evidence (the spectral anomaly score that TIMBRE computes for each client in each round) at three heterogeneity levels; the conclusion is also stated in Section 3.2. The experiments use CIFAR-10 × ResNet18; the attack lasts 50 rounds, and each round has 2 malicious and 8 benign clients among its 10 participants; we treat one participation of one client in one round as one sample, so each attack and heterogeneity level has 100 malicious and 400 benign participations. Each of the two thresholds corresponds to one kind of processing: a client whose current-round evidence exceeds the coordinate-processing threshold undergoes coordinate filtering in that round (only coordinates with same-sign support from other clients are kept), and a client whose evidence exceeds the strong-processing threshold enters the strong tier (its aggregation weight is multiplied by ms=0.01m_s=0.01). TPR and FPR are the percentages of malicious and benign participations whose evidence exceeds the corresponding threshold, so the FPR is exactly the fraction of benign participations that receive the corresponding processing because of their current-round evidence; the last two columns are the median evidence, with the 10% and 90% quantiles in brackets:

αAttackAUROC [95% CI]TPR: coordinateTPR: strongFPR: coordinateFPR: strongMalicious evidenceBenign evidence
0.2BadNets0.972 [0.954, 0.987]76542.000.755.22 [0.93, 10.99]−0.20 [−1.00, 0.83]
Batman0.921 [0.893, 0.944]36213.002.252.11 [0.62, 6.54]−0.19 [−0.98, 0.94]
Neurotoxin0.963 [0.946, 0.980]62381.251.003.85 [1.04, 8.89]−0.19 [−0.93, 0.77]
DBA0.954 [0.928, 0.975]65541.751.254.82 [0.83, 14.69]−0.19 [−0.87, 1.06]
0.5BadNets0.992 [0.986, 0.997]91781.250.757.96 [2.97, 14.04]−0.23 [−0.97, 0.84]
Batman0.971 [0.958, 0.982]65423.252.503.82 [1.41, 8.92]−0.26 [−0.83, 0.87]
Neurotoxin0.996 [0.991, 0.999]98831.501.007.97 [4.02, 15.33]−0.22 [−0.90, 0.93]
DBA0.980 [0.967, 0.989]83742.001.509.05 [1.88, 23.08]−0.18 [−0.90, 0.82]
1.0BadNets0.987 [0.981, 0.994]98935.753.259.54 [5.27, 18.58]−0.18 [−0.92, 0.97]
Batman0.973 [0.962, 0.983]76443.751.504.36 [2.03, 8.24]−0.18 [−0.86, 1.30]
Neurotoxin0.996 [0.991, 1.000]100982.751.0011.85 [6.96, 22.69]−0.20 [−0.91, 0.94]
DBA0.986 [0.979, 0.993]90834.252.509.50 [3.05, 24.09]−0.24 [−0.85, 1.07]
AuthorsOpenReview ↗

Response to Reviewer JBmo (2/8)

W1 / Q1 (continued). The AUROC is at least 0.921 at all three levels. The distribution of benign evidence hardly changes with heterogeneity because of how the score is computed: each client's score is standardized against the other clients of the same round, that is, the median of the round's scores is subtracted and the result is divided by their spread (MAD), and this spread grows as the data become more heterogeneous; the fraction of benign participations exceeding the coordinate-processing threshold does not rise with heterogeneity either. When heterogeneity increases, the change is mainly on the malicious side: at α = 0.2, the median malicious evidence drops to 2.11–5.22, and fewer malicious participations exceed the thresholds. TIMBRE does not need to identify the attacker in every round: the evidence of malicious participations that do not exceed the thresholds still enters the evidence accumulated across rounds (called debt in the main text), so a persistently poisoning client is processed as rounds accumulate. For example, under Batman at α = 0.2, only 36% of the malicious participations exceed the coordinate-processing threshold in a single round, yet the final ASR is 0.42%, rising to 88.89% without evidence accumulation (Appendix D.5).

Regarding the benign clients with highly skewed class distributions that you mentioned, we also split the benign participations into four groups by the quartiles of the client's largest class share, from group 1 (most balanced) to group 4 (most skewed). If skewed benign clients looked more like attackers, the Spearman rank correlation ρ\rho between the largest class share and the evidence should be positive and statistically significant, and the fraction exceeding the coordinate-processing threshold should increase from group 1 to group 4. The fractions (%) exceeding the coordinate-processing threshold in each group are:

αAttackSpearman ρ\rho (pp)Group 1 (most balanced)Group 2Group 3Group 4 (most skewed)
0.2BadNets0.06 (0.24)0.90.90.06.0
Batman0.05 (0.31)2.70.90.97.0
Neurotoxin−0.02 (0.74)0.90.00.04.0
DBA0.05 (0.31)0.00.91.94.0
0.5BadNets−0.00 (0.99)2.61.00.91.0
Batman−0.02 (0.71)7.91.90.92.9
Neurotoxin−0.04 (0.47)3.51.00.91.0
DBA−0.04 (0.43)3.51.90.91.9
1.0BadNets−0.06 (0.25)7.05.94.54.9
Batman0.02 (0.76)5.02.03.65.9
Neurotoxin−0.03 (0.57)3.05.02.72.0
DBA−0.06 (0.21)6.06.92.72.0

At all three heterogeneity levels, ρ\rho is not statistically significant (∣ρ∣≤0.06\lvert\rho\rvert\le0.06, p≥0.21p\ge0.21), and the fractions do not increase with skew; only at α = 0.2 is the most skewed group higher (4.0%–7.0%, versus 0–2.7% for the other three groups), still far below the 36%–76% of malicious participations in the same setting (the "TPR: coordinate" column above). Overall, skewed benign clients are not more likely to be misjudged.

AuthorsOpenReview ↗

Response to Reviewer JBmo (3/8)

W2 / Q2: MARS (NeurIPS 2025) has already used singular-value-related backdoor energy for malicious client detection. Although TIMBRE uses the full normalized spectral energy distribution, the paper does not explain its additional advantages over simpler spectral statistics such as the largest singular value, spectral entropy, or effective rank. Could the authors compare them within a unified detection framework?

The new Appendix D.6 replaces the per-matrix Hellinger distance with these three statistics in the same pipeline, keeping everything else unchanged and re-determining the thresholds with the same procedure; the results for the main settings are also in Table 2 of the main text. Worst final ASR (%) over four attacks (BadNets, Batman, Neurotoxin, and DBA) in the eight settings of Appendix D.3:

SettingTIMBRESpectral entropyEffective rankσ12\sigma_1^2 (unnormalized)FedAvg without defense
α = 0.5, f=2f=2 (default)1.381.782.0442.0489.03
α = 0.20.797.705.5078.4089.24
α = 0.30.720.471.4665.3494.09
α = 0.71.501.321.0160.5287.28
α = 1.00.991.100.740.5093.68
f=1f=11.531.281.191.4770.88
f=3f=30.5614.9633.1073.7289.82
f=4f=42.2376.5038.0190.7991.08

The unnormalized σ12\sigma_1^2 exceeds 40% in six of the eight settings, which is exactly the difference between TIMBRE and MARS: TIMBRE reads how the energy is distributed across components rather than how large it is. Spectral entropy and effective rank are also based on the normalized spectrum; they are close to TIMBRE in the default setting but rise markedly at α = 0.2 and with 3–4 malicious clients per round. For detection, the AUROC of each evidence in separating malicious from benign participations in the default setting and at α = 0.2 is:

SettingAttackTIMBRESpectral entropyEffective rankσ12\sigma_1^2 (unnormalized)
α = 0.5 (default)BadNets0.9920.9770.9600.809
Batman0.9710.9370.8610.973
Neurotoxin0.9960.9900.9830.631
DBA0.9800.9560.9340.818
α = 0.2BadNets0.9720.9560.9100.565
Batman0.9210.7480.7370.435
Neurotoxin0.9630.9150.8960.563
DBA0.9540.9580.9450.483

In the default setting, the three evidences based on the normalized spectrum all reach a high AUROC (0.861–0.996), and σ12\sigma_1^2 is comparable to TIMBRE only under Batman. At α = 0.2, σ12\sigma_1^2 falls to 0.435–0.565 under the four attacks, comparable to the chance level of 0.5; spectral entropy and effective rank fall to 0.748 and 0.737 under Batman, with final ASRs of 7.70% and 5.50%, whereas TIMBRE has an AUROC of 0.921 and a final ASR of 0.42%. The gap widens with 3–4 malicious clients per round: under Batman (3 or 4 malicious clients per round) and Neurotoxin (4 per round), the AUROC of spectral entropy and effective rank is only 0.618–0.705 and their final ASR is 14.96%–76.50%, while TIMBRE does not exceed 2.23% in these three cases. This is because spectral entropy and effective rank each summarize the whole distribution by one number, so distributions of different shapes may receive similar values, whereas the Hellinger distance compares the distributions themselves and simultaneously bounds the differences in the cumulative energy shares at every rank (Theorem A.2), without having to choose a particular summary statistic in advance.

Reviewer_YA17OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TIMBRE, a server-side backdoor defense for federated learning based on layer-wise normalized spectral energy distributions. It combines spectral anomaly detection, cross-round evidence accumulation, and coordinate-wise update processing to suppress malicious updates while preserving model utility. Experiments demonstrate strong defense performance across multiple attacks, datasets, and model architectures. However, the limited justification of its core spectral assumption, the reliance on pretrained models, and the absence of computational cost analysis leave important questions about its generalizability and practicality.

Critical strengths
  1. The paper introduces an interesting perspective for backdoor detection by analyzing the spectral energy distributions of client updates beyond conventional norm-based and direction-based statistics.
  2. The proposed method operates entirely on the server without modifying client training or requiring trusted reference data during deployment.
  3. The experiments demonstrate promising backdoor suppression across multiple attacks, datasets, and heterogeneous settings.
Critical weaknesses
  1. Insufficient justification for the spectral assumption. Spectral energy concentration is not unique to backdoor updates. Benign clients with highly non-IID or class-skewed data may exhibit similar patterns, while malicious updates may not necessarily have concentrated spectra. Figure 1 examines only one convolutional layer under BadNets, and the theoretical analysis in Appendix A does not establish a direct connection between spectral anomalies and backdoor behavior.
  2. Limited evaluation across training stages. The experiments primarily start from extensively pretrained clean models. For example, CIFAR-10 with ResNet18 uses a model pretrained for 2,000 rounds before launching attacks. It remains unclear whether TIMBRE is equally effective during early-stage federated training, when benign updates may exhibit different spectral characteristics.
  3. Missing computational overhead analysis. TIMBRE performs layer-wise SVD and coordinate-wise filtering on client updates, but the paper does not report the associated computational cost, making its practical efficiency and scalability difficult to assess.
Questions for authors
  1. Could the authors further discuss how TIMBRE distinguishes backdoor-related spectral anomalies from similar patterns caused by highly non-IID or class-skewed benign clients?
  2. Could the authors discuss whether backdoor updates with more distributed spectral energy could evade the proposed detection criterion, and provide further theoretical justification for the connection between spectral anomalies and backdoor behavior?
  3. How does TIMBRE perform when federated training starts from random initialization rather than a pretrained clean model?
  4. A runtime comparison with FedAvg, MARS, and other relevant defenses would help clarify the additional computational overhead and scalability of TIMBRE. 5.Although the paper discusses the differences between TIMBRE and existing spectral-based methods such as MARS, could the authors further clarify the advantages of normalized spectral energy distributions over existing spectral features, ideally through a direct comparison of their detection capabilities?
8 comments
AuthorsOpenReview ↗

Response to Reviewer YA17 (1/5)

We thank the reviewer for the positive assessment and the comments.

W1 / Q1: The spectral assumption is insufficiently justified: energy concentration is not unique to backdoor updates; benign clients with highly non-IID or class-skewed data may exhibit similar patterns, and malicious updates may not have concentrated spectra; Figure 1 examines only one convolutional layer under BadNets, and the theory in Appendix A does not establish a direct connection between spectral anomalies and backdoor behavior. How does TIMBRE distinguish backdoor-related spectral anomalies from similar patterns of such benign clients?

The new Appendix D.4 examines this with the current-round evidence actually used by the server during the runs; the conclusion is also stated in Section 3.2. The current-round evidence is the anomaly score that TIMBRE computes for each client in each round. It is aggregated from layer-wise scores on 17 update matrices (the 8 highest are used), rather than taken from the single layer of Figure 1; when it exceeds the coordinate-processing threshold (2.91 on CIFAR-10 × ResNet18), the client undergoes coordinate filtering in that round. We ran two checks at α = 0.2, 0.5, and 1.0 under four attacks (BadNets, Batman, Neurotoxin, and DBA). The attack lasts 50 rounds, and each round has 2 malicious and 8 benign clients among its 10 participants; we treat one participation of one client in one round as one sample, so each attack and heterogeneity level has 100 malicious and 400 benign participations.

First, if more heterogeneous data made benign clients more likely to be taken as anomalous, the benign scores should rise overall as heterogeneity increases. The actual values are (ranges over the four attacks):

αMedian benign score90% quantile of benign scoresBenign participations exceeding the coordinate-processing threshold (%)
0.2−0.20 to −0.190.77–1.061.25–3.00
0.5−0.26 to −0.180.82–0.931.25–3.25
1.0−0.24 to −0.180.94–1.302.75–5.75

At all three levels, the benign scores concentrate around 0, their 90% quantile does not exceed 1.30, far below 2.91, and the fraction exceeding the threshold does not rise as heterogeneity increases. The reason lies in how the score is computed: each client's score is standardized against the other clients of the same round, that is, the median of the round's scores is subtracted and the result is divided by their spread (MAD). The more heterogeneous the data, the more benign updates differ from one another and the larger this spread becomes, so the benign scores do not rise overall.

AuthorsOpenReview ↗

Response to Reviewer YA17 (2/5)

W1 / Q1 (continued). Second, we split the benign participations into four groups by the quartiles of the client's largest class share, from group 1 (most balanced) to group 4 (most skewed). If skewed benign clients looked more like attackers, the Spearman rank correlation ρ\rho between the largest class share and the evidence should be positive and statistically significant, and the fraction exceeding the coordinate-processing threshold should increase from group 1 to group 4. The fractions (%) exceeding the coordinate-processing threshold in each group are:

αAttackSpearman ρ\rho (pp)Group 1 (most balanced)Group 2Group 3Group 4 (most skewed)
0.2BadNets0.06 (0.24)0.90.90.06.0
Batman0.05 (0.31)2.70.90.97.0
Neurotoxin−0.02 (0.74)0.90.00.04.0
DBA0.05 (0.31)0.00.91.94.0
0.5BadNets−0.00 (0.99)2.61.00.91.0
Batman−0.02 (0.71)7.91.90.92.9
Neurotoxin−0.04 (0.47)3.51.00.91.0
DBA−0.04 (0.43)3.51.90.91.9
1.0BadNets−0.06 (0.25)7.05.94.54.9
Batman0.02 (0.76)5.02.03.65.9
Neurotoxin−0.03 (0.57)3.05.02.72.0
DBA−0.06 (0.21)6.06.92.72.0

At all three heterogeneity levels, ρ\rho is not statistically significant (∣ρ∣≤0.06\lvert\rho\rvert\le0.06, p≥0.21p\ge0.21), and the fractions do not increase with skew; only at α = 0.2 is the most skewed group higher (4.0%–7.0%), still far below the 36%–76% of malicious participations in the same setting. Overall, skewed benign clients are not more likely to be misjudged.

The connection between spectral anomalies and backdoor behavior can be viewed in two steps. The step from "concentrated energy" to "flagged" is given by the new Theorem A.9: if the energy of an update is markedly more concentrated than the reference and this concentration exceeds the normal spread among benign clients, its current-round anomaly score must exceed the threshold. For the step from "backdoor training" to "concentrated energy", the new Proposition A.3 gives a sufficient condition (see our response to Q2 below); whether this condition holds for actual updates is tested by the experiment in Appendix D.4: with the actual score introduced above (aggregated over many matrices, not only the single layer of Figure 1), the AUROC in separating malicious from benign participations is at least 0.921 under four attacks and three heterogeneity levels, showing that the spectral distributions of backdoor updates are indeed distinguishable from those of benign updates in practice.

AuthorsOpenReview ↗

Response to Reviewer YA17 (3/5)

W2 / Q3: The experiments mainly start from extensively pretrained clean models (e.g., CIFAR-10 with ResNet18 is pretrained for 2,000 rounds before the attacks), while benign updates in early-stage training may have different spectral characteristics. How does TIMBRE perform when federated training starts from random initialization rather than a pretrained clean model?

The new Appendix D.10 runs TIMBRE and FedAvg on CIFAR-10 × ResNet18 with α = 0.5 and 2 malicious clients per round, starting from random initialization and from the FedAvg checkpoints at rounds 50, 200, and 600. The attacks are BadNets and Neurotoxin, representing basic backdoor injection and an attack designed for backdoor durability. Training runs for 200 rounds from random initialization and for 50 rounds from the other starting points; all starting points keep the same thresholds determined on a near-converged model, without re-determining them for a new starting point, and the main results are also stated in Section 5.2. Final results (%):

Start (ACC at start)RoundsBadNets ASR: TIMBRE / FedAvgNeurotoxin ASR: TIMBRE / FedAvgNo-attack ACC: TIMBRE / FedAvg
Random init. (about 10)2002.50 / 90.691.40 / 68.3468.48 / 79.35
Round 50 (44.9)504.96 / 63.330.89 / 40.4857.62 / 63.18
Round 200 (75.1)502.39 / 81.221.17 / 81.6278.76 / 79.55
Round 600 (87.6)501.56 / 85.634.16 / 91.1386.83 / 88.29
Round 2000 (91.6)501.20 / 76.881.38 / 83.0690.61 / 92.06

TIMBRE suppresses the backdoor from every starting point. The separation by the current-round evidence is weaker only at the very beginning of training: in the first 50 rounds from random initialization, the AUROC under BadNets and Neurotoxin is 0.75–0.78, rising to 0.95–1.00 in rounds 51–200; from the checkpoints at rounds 50, 200, and 600 it is 0.87–1.00.

The accuracy cost is concentrated in the stage far from convergence. To see whether this cost vanishes with training, we extended the no-attack run from random initialization to 500 rounds; ACC (%):

Round100200300400500
TIMBRE54.8768.4875.5480.2983.07
FedAvg63.8079.4284.7786.8687.75

The gap narrows from about 11 percentage points at round 200 to about 5 at round 500, but does not vanish within 500 rounds.

W3 / Q4: The computational cost of layer-wise SVD and coordinate-wise filtering is not reported, making practical efficiency and scalability hard to assess; a runtime comparison with FedAvg, MARS, and other defenses would help.

The new Appendix D.13 measures the server aggregation time per round on the same NVIDIA L20: CIFAR-10 × ResNet18, 10 clients per round, no attack, median over rounds 2002–2020, with no other training job on the node during timing. Results (s):

FedAvgAlignInsRLRFoolsGoldMulti-KrumMARSFLAMEMulti-MetricsTIMBREDeepSight
0.030.050.060.230.240.370.420.881.3012.31

TIMBRE takes 1.30 s per round, between Multi-Metrics and DeepSight. Of this, the spectral scoring itself takes about 0.02 s per round and coordinate processing about 0.26 s; what grows with the number of clients is mainly the singular value decomposition, which is performed separately for each client, so its time is proportional to the number of clients per round. Measured on random updates with the layer shapes of ResNet18, the singular value decompositions take 0.51, 1.01, 2.54, and 5.20 s for 10, 20, 50, and 100 clients per round.

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

TIMBRE is a server-side defense against backdoors in heterogeneous federated learning. TIMBRE computes a layer-wise normalized spectral energy distribution for each uploaded update via SVD, compares it with a same-round reference spectrum from other clients, converts the deviation into a robust anomaly score, and accumulates this score across rounds as a debt. Based on current-round evidence and accumulated debt, clients are assigned to tiers; suspicious updates are norm-clipped, coordinate-filtered using same-sign support from non-target clients, interval-clipped, and down-weighted during aggregation.

Critical strengths
  1. the manuscript is easy to follow and well-organized.
  2. The layer-wise spectral energy perspective is a good evidence source for federated backdoor detection. It is scale- and direction-invariant to some extent, and the manuscript provides a useful analysis linking normalized spectra to low-rank approximation and a Hellinger-distance lower bound
Critical weaknesses
  1. The assumption might not be hold to some extent. TIMBRE requires the server to inspect every individual client update layer by layer, and its reference statistics assume that benign clients form the majority in every round. The paper acknowledges this limitation, however, with some secure aggregation strategies, the client updates might not be hidden. The claim of being a general server-side defense is therefore not suitable.
  2. The manuscript only test the TIMBRE on some simple classification tasks, more complex tasks should be considered. such as NLP tasks.
  3. The theoretical analysis only bounds spectral distance but does not establish detection guarantees or false-positive control under heterogeneous benign updates.
Questions for authors
  1. How does TIMBRE behave under secure aggregation or a privacy-preserving server that cannot inspect client updates?
  2. Please provide more complex backdoor tasks.
  3. Please polish the provided theoretical analysis.
3 comments
AuthorsOpenReview ↗

Response to Reviewer FpSW (1/2)

We thank the reviewer for the comments and for the positive remarks on the organization of the paper and the spectral-evidence perspective.

W1 / Q1: TIMBRE inspects every client update layer by layer, and its reference statistics assume a benign majority in every round; under some secure aggregation schemes client updates are hidden, so the claim of a general server-side defense is not suitable. How does TIMBRE behave under secure aggregation or a privacy-preserving server that cannot inspect client updates?

TIMBRE is a server-side defense for the setting in which the server can read per-client updates: the abstract states that its evidence is taken from per-client updates, and the defender model in Section 2.2 assumes that the server can read the updates uploaded by the clients and that fewer than half of the participants in each round are malicious. The defenses we compare with (Multi-Krum, FLAME, DeepSight, MARS, AlignIns, and others) likewise need to see per-client updates. Therefore, under secure aggregation that hides individual updates, TIMBRE in its current form cannot be applied directly, and Section 7 lists this as a limitation.

W2 / Q2: TIMBRE is tested only on simple classification tasks; more complex tasks such as NLP tasks should be considered. Please provide more complex backdoor tasks.

Besides CIFAR-10 and CIFAR-100, the original submission already included the 200-class Tiny-ImageNet (Appendix D.1) and a text experiment with a four-layer LSTM on IMDB (Appendix D.6 of the original submission, D.2 in the revision). The revision adds a pretrained ALBERT (albert-base-v2) to the same section. TIMBRE analyzes the spectra of 10 weight matrices of ALBERT one by one, namely the word embedding, the position embedding, the embedding projection, the Q/K/V/O projections of attention, the two feed-forward layers, and the pooler, and aggregates the 5 highest scores, a proportion comparable to that of ResNet18 (8 of 17); the thresholds are calibrated with the same rule as in the image experiments. Of the two attacks, Word-trigger is the text counterpart of BadNets, and Neurotoxin was itself evaluated on IMDB in its original paper. Results (%) after 50 rounds of continued training on IMDB:

ModelConditionFedAvg ACCFedAvg ASRTIMBRE ACCTIMBRE ASR
LSTMWord-trigger79.27100.0075.0414.10
Neurotoxin78.56100.0076.6016.08
No attack78.29—76.68—
ALBERTWord-trigger91.88100.0090.374.88
Neurotoxin91.89100.0091.025.49
No attack91.92—91.71—

Under TIMBRE, the ASR is no higher than the rate at which the trigger word alone flips negative reviews for an unpoisoned model (18.80% for the LSTM and 10.26% for ALBERT); without attack, the ACC is 1.61 (LSTM) and 0.21 (ALBERT) percentage points below FedAvg.

AuthorsOpenReview ↗

Response to Reviewer FpSW (2/2)

W3 / Q3: The theoretical analysis only bounds the spectral distance and establishes neither detection guarantees nor false-positive control under heterogeneous benign updates. Please polish the theoretical analysis.

The revision adds to Appendix A the detection guarantees and false-positive control you asked for: Proposition A.3 and results A.5–A.10. The former explains under which condition backdoor training concentrates the spectral energy of an update, and the latter connect the existing lower bound on the spectral distance to the score actually used in the main text. The main results are also stated in the main text (Proposition 1 in Section 3.2 and Theorem 1 in Section 4.2). All of these results concern a given round: given the updates of the clients in that round, which clients must exceed the threshold, and at most how many benign clients can be flagged.

  • Detection guarantee (Proposition A.3, Theorem A.9). Proposition A.3 explains when backdoor training concentrates the energy: the gradients of poisoned samples accumulate coherently along the trigger direction, and as long as this component is stronger than the rest of the update, the energy concentrates in the leading spectral component. Theorem A.9 then shows that if, on enough update matrices, a malicious client concentrates its energy in a few spectral components markedly more than the reference spectrum of the other clients of the round does, and this concentration exceeds the normal spread of the distances among its benign peers, then its anomaly score in this round (the current-round evidence in the main text) must exceed the threshold. The quantities in the condition that involve other clients are taken from benign clients only, so the result holds no matter how the other malicious clients construct their updates.
  • False-positive control (Theorem A.6). The current-round evidence is a score standardized against the participants of the same round: the median of the round's scores is subtracted and the result is divided by their spread (MAD). As long as benign clients form the majority, both quantities are sandwiched by order statistics of the benign clients' scores (Lemma A.5), and malicious clients cannot push them out of this range whatever values they take. In other words, malicious clients cannot push a benign client over the threshold by manipulating this standardization: a benign client is flagged only if its score is high among the benign clients of the round. On this basis, Theorem A.6 gives lower and upper bounds on the number of flagged benign clients in the round, and the bounds depend only on the actual scores of the benign clients in the round (the spectra of malicious clients can still affect these scores through the reference spectra, as noted after Theorem A.6 in Appendix A); the data distributions of benign clients may differ and their scores may be dependent, so the result applies to heterogeneous benign updates.
  • False-positive rate and threshold calibration (Proposition A.10). The bounds of Theorem A.6 vary with the actual distribution of the benign scores in the round and are not a fixed rate that holds for any data; Proposition A.10 shows, with a round consisting only of benign clients, that such a fixed rate cannot be derived from a benign majority alone. In other words, how many false positives occur depends on the distribution of benign scores on the given data, so the thresholds must be set after observing benign scores on the corresponding dataset and model. The calibration in Appendix C.4 does exactly this: on one no-attack run, the thresholds are chosen so that the fractions of benign clients that trigger coordinate processing and strong down-weighting through their current-round scores do not exceed 10% and 5%; the revision states these two fractions explicitly as constraints on the calibration run.

If these additions address your main concerns, we would be grateful if you would consider re-evaluating the paper; we are happy to clarify any remaining questions during the discussion period.

AuthorsOpenReview ↗

Kind reminder

Dear Reviewer FpSW,

Thank you again for your review. As the discussion period ends on November 18 (AoE), we would like to check whether our responses address your concerns. In brief, the revision adds a pretrained ALBERT text experiment (Appendix D.2), adds detection guarantees and bounds on flagged benign clients to Appendix A (Proposition A.3 and results A.5–A.10; Proposition 1 and Theorem 1 in the main text), and clarifies that TIMBRE, like the defenses we compare with, targets the setting in which the server reads per-client updates. If anything remains unclear, we would be glad to clarify before the deadline.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the three reviewers for their comments. TIMBRE is a federated backdoor defense that runs entirely on the server: it reads how each update allocates its spectral energy within each layer, rather than only its magnitude or direction, and imposes graded restrictions according to the current-round strength and the cross-round persistence of the evidence, without issuing client-level exclusion verdicts. All three reviewers found the paper clear and easy to follow and appreciated the layer-wise spectral energy perspective as a source of evidence (FpSW, YA17, JBmo); YA17 and JBmo also noted that TIMBRE is easy to deploy, since it runs entirely on the server and requires no change to local training.

The main concerns fall into four groups, which the new results in the revision address as follows:

  • Is the spectral evidence still discriminative under strong heterogeneity? (YA17, JBmo) At α = 0.2, 0.5, and 1.0, the AUROC of the current-round evidence in separating malicious from benign participations is at least 0.921, and the evidence of benign participations hardly changes with heterogeneity or label skew (Section 3.2; Appendix D.4).
  • Does the effect come from the spectral evidence or from the strong down-weighting? (JBmo) Keeping the response mechanism and replacing only the evidence with random numbers leaves the final ASR at 73%–101% of that of FedAvg without defense in the same setting; replacing it with the largest singular-value energy fails to suppress the backdoor in most settings, and replacing it with spectral entropy or effective rank becomes markedly worse under stronger heterogeneity or more malicious clients (Table 2; Appendix D.6).
  • Can the theory give detection and false-positive control? (FpSW) We add a sufficient condition under which a malicious client exceeds the threshold in a round, bounds on the number of flagged benign clients, and a condition under which backdoor training concentrates the spectral energy (Proposition 1 in Section 3.2 and Theorem 1 in Section 4.2; Appendix A).
  • Do the thresholds have to be tuned per setting? (JBmo) Each dataset–model combination uses a single set of thresholds for all attacks, heterogeneity levels, and training starting points; applying the CIFAR-10 × ResNet18 thresholds unchanged to CIFAR-100 × ResNet18 changes the ASR by less than 0.5 percentage points relative to its own thresholds (Sections 2.2 and 5.2; Appendices D.3, D.10, and D.11).

We have uploaded a revised PDF in which new and changed text is shown in blue. Using the 10-page limit of the discussion phase, the main text now includes Proposition 1 and the detection results under heterogeneity in Section 3.2; Theorem 1 in Section 4.2; in Section 5.2, the replacement of the scoring statistic (Table 2), the training stage and threshold transfer, as well as ALBERT, the random-target control, and the joint adaptive attacks; and, in Section 2.2, the statement that the thresholds are determined once before deployment. Everything else is in the appendix. The revision reorders Appendix D, and all appendix numbers below refer to the revision. The other new experiments are as follows; details are in the individual responses:

  • Text task (Appendix D.2; FpSW W2): we add a pretrained ALBERT; under two attacks, the ASR drops from 100% without defense to 4.88% and 5.49%, and the no-attack ACC is 0.21 percentage points lower.
  • Joint adaptive attacks (Appendix D.7; JBmo W5): on top of spectral mimicry, the attacker also evades same-sign coordinate support and cross-round accumulation; over ten configurations, the final ASR of TIMBRE is 0.46%–5.80%, and that of FedAvg without defense is 4.91%–73.89%.
  • Training from random initialization (Appendix D.10; YA17 W2, JBmo W7): five training starting points use the same thresholds, and the final ASR is 0.89%–4.96%; early training has an accuracy cost: after 500 rounds from random initialization, the no-attack ACC is still about 5 percentage points below FedAvg.
  • Hyperparameter sensitivity (Appendix D.12; JBmo W6): over 37 values of the six parameters named by the reviewer, the final ASR is 0.20%–2.47%; when the evidence is weak, the value of the coordinate-tier weight multiplier mcm_c matters more.
  • Runtime (Appendix D.13; YA17 W3): TIMBRE takes 1.30 s per round for server-side aggregation, compared with 0.03, 0.37, and 12.31 s for FedAvg, MARS, and DeepSight; the main cost is the per-client singular value decomposition.

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.