acceptodds
Under review as a conference paper at ICLR 2027

FedRASR: Risk-Aware Soft Reweighting against Persistent Poisoning in Federated Load Forecasting

Abstract

Federated learning enables collaborative load forecasting without pooling private smart-meter data, but persistent poisoning is difficult to detect because small malicious updates accumulate across rounds and are obscured by regional non-IID variation. To evaluate this threat, we propose Hist-Residual, which repeatedly perturbs profile-aware targets, and Profile-LIE, which additionally shapes uploads using benign round-level coordinate statistics. Existing robust aggregation rules rely on current-round geometry, and cross-round detectors may confuse regional heterogeneity with malicious temporal behavior. To address this, we develop a Federated Risk-Aware Soft Reweighting algorithm (FedRASR), a five-stage server-side defense that forms compact update features, constructs region-calibrated residual trajectories, derives FFT-based frequency evidence, estimates risk, and softly reweights the complete updates. Under the evaluated threat setting, experiments on four regional smart-meter collections with 1,200 clients show that FedRASR attains the lowest RMSE and MAE among the compared methods under both attacks, remains competitive under clean training, and benefits from both frequency scoring and soft reweighting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
federated learning, persistent poisoning, robust aggregation, non-IID data, time-series forecasting, smart grid

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VgzoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates persistent poisoning attacks in federated load forecasting under non-IID data distributions. To mitigate the two proposed poisoning attacks, the paper proposes FedRASR, a server-side defense that constructs region-calibrated client update trajectories, extracts temporal frequency characteristics using FFT, estimates client risk, and adaptively reweights model updates during aggregation. The results show that FedRASR achieves lower forecasting errors than the evaluated robust aggregation baselines.

Critical strengths
  1. The paper addresses a meaningful problem in heterogeneous federated learning.

  2. The evaluation includes multiple regional datasets and ablations.

Critical weaknesses
  1. The evaluated attacks may exhibit temporal signatures that are particularly favorable to the proposed frequency-domain detector. The authors should evaluate additional poisoning schedules, such as randomized and adaptive attacks, to establish whether the proposed method captures general persistent malicious behavior.

  2. FedRASR relies on regional medians and MAD-based thresholds to separate benign heterogeneity from malicious behavior. However, the threat model specifies only a global malicious-client fraction and does not impose a corresponding bound on the malicious fraction within each region. Attackers could therefore concentrate compromised clients in particular regions and potentially bias both the reference centers and risk thresholds.

  3. Why the results in Tables S7–S8 show substantial regional variation?

  4. The paper discusses FreqFed, FedRDF, and FLDetector but does not include them in the experimental evaluation. The compared methods are relatively outdated.

Questions for authors

Please check weaknesses.

3 comments
AuthorsOpenReview ↗

Response by Authors (Part 1)

We are grateful to the reviewer for recognizing the importance of the problem and the evaluation across regional datasets. We will clarify the attack schedules, reference conditions, regional variation, and baseline comparisons separately below.

W1: Robustness beyond alternating schedules

Thanks for pointing it out. We agree that evaluating only alternating attacks limits the generalizability of FedRASR. Therefore, we additionally evaluate constant-sign and randomized-sign attacks, and compare zero-filled and observed-only trajectories. As shown in Table 1, zero-filled FedRASR achieves the lowest mean RMSE and MAE among the three compared methods across all six settings, demonstrating that its advantage is not restricted to alternating attacks.

Table 1. Attack schedule and trajectory comparisons.

AttackSignFedAvg RMSE / MAE ↓FedRASR Zero-filled RMSE / MAE ↓FedRASR Observed-only RMSE / MAE ↓
Hist-ResidualAlternating0.1037 ± 0.0024 / 0.0807 ± 0.00190.0937 ± 0.0018 / 0.0683 ± 0.00220.1009 ± 0.0017 / 0.0776 ± 0.0010
Hist-ResidualConstant0.0829 ± 0.0015 / 0.0517 ± 0.00100.0828 ± 0.0018 / 0.0514 ± 0.00130.0831 ± 0.0014 / 0.0515 ± 0.0009
Hist-ResidualRandomized0.0947 ± 0.0104 / 0.0690 ± 0.01460.0945 ± 0.0083 / 0.0689 ± 0.01140.0948 ± 0.0092 / 0.0695 ± 0.0128
Profile-LIEAlternating0.1031 ± 0.0022 / 0.0787 ± 0.00170.0940 ± 0.0021 / 0.0683 ± 0.00210.1008 ± 0.0017 / 0.0764 ± 0.0013
Profile-LIEConstant0.0837 ± 0.0018 / 0.0525 ± 0.00140.0834 ± 0.0018 / 0.0520 ± 0.00150.0839 ± 0.0016 / 0.0522 ± 0.0012
Profile-LIERandomized0.0948 ± 0.0107 / 0.0686 ± 0.01430.0945 ± 0.0084 / 0.0685 ± 0.01110.0946 ± 0.0092 / 0.0686 ± 0.0124

W2: Concentrated Contamination

We thank the reviewer for raising this concern. Our threat model is based on a fixed malicious-client pool randomly sampled from the global population, with a 30% malicious-client ratio (to satisfy the minority assumption for adversaries), following the commonly adopted benign-majority setting in robust federated learning [1, 2]. On the other hand, only a subset of clients participate in the training (normally no more than 20% of all clients). The probability that malicious clients dominate a particular region's participating clients is very low. Thus, the scenario (attackers dominate a specific region) is unlikely under our considered threat model and represents a more extreme adversarial setting.

Nevertheless, to directly address this concern, we further evaluate a challenging scenario where 49% of clients in a specific region (Portugal) are malicious (Must < 50% to follow the protocol). This setting substantially increases regional adversarial concentration compared with our original random-compromise setting. In the tables below, FedRASR still works and achieves the lowest mean RMSE and MAE among the three compared methods and these results demonstrate FedRASR's robustness under concentrated regional poisoning.

Hist-Residual

RegionRMSE FedRASR ↓RMSE FedAvg ↓RMSE FreqFed ↓MAE FedRASR ↓MAE FedAvg ↓MAE FreqFed ↓
Portugal0.129340.159130.137010.112570.138720.12754
Ausgrid0.083550.084780.084250.061500.074130.06378
London0.087750.092830.090930.062380.076790.07254
Swiss0.103200.107620.104280.070320.090370.08461
Four-region macro0.100960.111090.104120.076690.095000.08712

Profile-LIE

RegionRMSE FedRASR ↓RMSE FedAvg ↓RMSE FreqFed ↓MAE FedRASR ↓MAE FedAvg ↓MAE FreqFed ↓
Portugal0.132740.160620.147290.110310.141830.12015
Ausgrid0.081760.084300.083930.062140.073280.06851
London0.088020.090410.089650.063920.078970.07253
Swiss0.096510.104130.098730.072160.084460.07882
Four-region macro0.099760.109870.104900.077130.094640.08500

[1] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. Dba: Distributed backdoor attacks against federated learning. In International conference on learning representations. 2019

[2] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International conference on machine learning. 2018

AuthorsOpenReview ↗

Response by Authors (Part 2)

W3: Regional Variation

Thanks for pointing it out. It is the regional variation that reflects the heterogeneous (non-i.i.d.) data among clients in different regions. This is actually one of our challenges and motivations in Attack and Defense in Federated Load Forecasting. To address this, we develop regional calibration to mitigate the impact of benign regional heterogeneity on poisoning detection. On the other hand, under clean training, the performance differences are already observed ( clean FedAvg in Table 3). To further demonstrate the effectiveness of FedRASR across different regions, we conduct experiments and expand the testing clients from 10 to 300 per region (this will further increase the heterogeneity). As shown in Table 3, FedRASR consistently outperforms FedAvg across all four regions under both attacks, achieving 7.7%–19.1% RMSE reductions under Hist-Residual and 4.8%–19.4% under Profile-LIE. These results demonstrate that FedRASR's improvements are not limited to a particular region.

Table 3. Regional RMSE on the expanded held-out test cohort.

RegionClean FedAvgHist-Residual FedAvgHist-Residual FedRASRProfile-LIE FedAvgProfile-LIE FedRASR
Portugal0.11180.15790.12770.15900.1281
Ausgrid0.07400.08330.07620.07940.0752
London0.08030.08920.08230.08550.0814
Swiss0.09270.10550.09600.10170.0951

W4: More baselines.

Thanks for pointing it out. We have added three baselines (FreqFed, FedRDF, and FLDetector) as requested. The results are in the table below.

MethodRMSE Clean ↓RMSE Hist-Residual ↓RMSE Profile-LIE ↓MAE Clean ↓MAE Hist-Residual ↓MAE Profile-LIE ↓
FedRASR (ours)0.0897 ± 0.00220.0955 ± 0.00220.0950 ± 0.00200.0628 ± 0.00170.0719 ± 0.00290.0703 ± 0.0024
FreqFed0.0989 ± 0.00780.1044 ± 0.00990.0956 ± 0.00430.0696 ± 0.00630.0769 ± 0.01190.0760 ± 0.0038
FLDetector0.0928 ± 0.00290.1039 ± 0.00320.1063 ± 0.00550.0646 ± 0.00260.0779 ± 0.00320.0798 ± 0.0052
FedRDF0.0931 ± 0.00290.1036 ± 0.00350.1056 ± 0.00470.0651 ± 0.00230.0776 ± 0.00340.0791 ± 0.0046

Dear Reviewer:

Thank you very much for your constructive comments. We hope that our responses have addressed your concerns. Please feel free to let us know if any further clarification is needed.

Best regards,

The authors

Reviewer_4HfmOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedRASR, a server-side defense against persistent poisoning attacks in federated load forecasting. It introduces two attacks, Hist-Residual and Profile-LIE, and develops a five-stage defense that combines compact update representation, regional residual calibration, FFT-based frequency analysis, risk estimation, and soft reweighting. Experiments on four regional smart-meter datasets involving 1,200 clients show improved forecasting performance under the evaluated attacks. However, several methodological assumptions and design choices require stronger justification, and the current evaluation does not fully establish the generality of the proposed defense.

Critical strengths
  1. The paper studies persistent poisoning in federated load forecasting, where regional Non-IID characteristics make malicious updates difficult to distinguish from legitimate variations.
  2. The method analyzes cross-round client behavior rather than relying exclusively on the geometry of individual-round updates.
  3. Regional residual calibration provides a reasonable way to reduce the influence of stable regional differences during anomaly detection.
  4. Soft reweighting avoids immediately excluding suspicious clients and preserves the possibility that their updates contain useful information.
  5. The experiments cover four regional datasets, 1,200 clients, multiple baselines, five random seeds, and ablation studies.
Critical weaknesses
  1. The risk score lacks a clear optimization objective or theoretical guarantee that it separates malicious updates from benign heterogeneity. Spectral statistics capture temporal patterns rather than maliciousness, while centering and normalization may obscure sustained offsets and attack magnitude.

  2. Zero-filling missing rounds confounds participation patterns with update behavior. Even constant benign residuals can generate artificial frequency components, making risk scores sensitive to client availability rather than maliciousness.

  3. Regional calibration assumes predominantly benign and stable reference sets, but authenticated membership does not ensure this. Regionally concentrated attackers, benign distribution shifts, and reference selection based on previous defense weights may distort thresholds and create feedback loops.

  4. Soft reweighting does not bound adversarial influence. Because weights remain positive and updates are not norm-bounded, an attacker can retain substantial influence by submitting sufficiently large updates, regardless of its risk ranking.

  5. The mechanism combines multiple heuristic components without adequately justifying their necessity or interactions. Since the evaluated attacks share an alternating pattern aligned with the spectral detector, broader ablations and sensitivity analyses are needed across participation patterns, regional contamination levels, and non-oscillatory or adaptive attacks.

Questions for authors
  1. Under what assumptions on benign dynamics, attack patterns, and client sampling can the spectral score distinguish malicious residuals from benign heterogeneity, particularly persistent directional bias from benign oscillation?

  2. Why are missing rounds zero-filled rather than explicitly masked? How does the method separate residual dynamics from participation-mask artifacts, and has it been compared with a mask-aware representation, including for constant residuals under random participation?

  3. How much information is lost through compact update representation, and can attacks concentrated in discarded update components evade detection?

  4. Does regional calibration require a benign majority or an explicit contamination bound within each region and reference set? How does it handle regionally concentrated attackers, benign distribution shifts, and feedback caused by selecting references using previous defense weights?

  5. Under what assumptions on update norms, sample counts, and weight separation does Eq. (15) bound malicious influence? Can the authors provide a conditional robustness result that explicitly accounts for the positive weight floor, or limit their claims to the evaluated attacks?

  6. How sensitive is FedRASR to frequency-score weights, regional thresholds, rank gates, and the minimum aggregation weight, and are these settings transferable across datasets? Please also report held-out test results and evaluate different malicious-client ratios, participation rates, attack patterns, and relevant frequency-based baselines.

6 comments
AuthorsOpenReview ↗

Response by Authors (Part 1)

We appreciate the reviewer’s recognition of the cross-round analysis, regional calibration, soft reweighting, and the breadth of the evaluation. We will clarify the roles of these components, the supporting evidence, and the conditions under which our conclusions apply.

W1 & Q1: Risk Score Design and Theoretical Motivation

The primary objective of FedRASR is to continuously adjust client contributions based on estimated risks, rather than making permanent binary classifications of benign and malicious clients. Risk score is used for soft reweighting rather than binary malicious-client classification. From an information-theoretic perspective, hard isolation has lower Shannon entropy and entirely discards the contributions of flagged clients, potentially losing useful information when benign clients are misclassified. In contrast, soft reweighting has higher Shannon entropy and preserves graded contributions from more clients, allowing the aggregation to retain a richer diversity of information while suppressing suspicious updates. This information-entropy-preservation principle is one of our key motivations for adopting risk-aware soft reweighting and is the theoretical foundation of FedRASR.

FedRASR integrates regional calibration, spectral characteristics, and directional evidence rather than relying solely on frequency statistics. Its effectiveness relies on three conditions: (i) regional references predominantly reflect representative benign behavior; (ii) malicious updates exhibit distinguishable deviations in retained spectral features or directional changes relative to benign references; and (iii) sufficient historical observations are available to characterize cross-round behavior. Please note that benign oscillations are evaluated relative to their regional references rather than automatically treated as malicious, while persistent attacks with abnormal temporal variations can still receive soft penalties.

We further validate the benefits of soft reweighting through the ablation experiments in Appendix Table S9. In the table below, compared with hard isolation, the full FedRASR reduces the normalized malicious aggregation share from 0.1426 to 0.1037 under Hist-Residual and from 0.1410 to 0.1035 under Profile-LIE, while maintaining average benign weights above 0.95.

Table 1. Aggregation mechanism comparison between Hard Isolation and FedRASR.

AttackMethodMalicious share ↓Malicious weight ↓Benign weight ↑
Hist-ResidualHard isolation0.1426 ± 0.04140.3862 ± 0.05140.9917 ± 0.0067
Hist-ResidualFedRASR0.1037 ± 0.03460.2573 ± 0.04010.9592 ± 0.0048
Profile-LIEHard isolation0.1410 ± 0.03890.3811 ± 0.04390.9905 ± 0.0067
Profile-LIEFedRASR0.1035 ± 0.03700.2565 ± 0.04480.9608 ± 0.0042

W2 & Q2: Zero-filling and Participation Mask Effects

Our motivation for zero-filling is to preserve the actual temporal intervals between client uploads. FedRASR constructs trajectories along global communication rounds, where missing uploads are represented by zeros. For observed-only encoding, it concatenates available updates, compressing the intervals between participation events. We acknowledge that participation masks can introduce additional frequency components. We therefore conduct both synthetic and end-to-end experiments to examine their effects.

Synthetic mask experiments. We fix the underlying residual signals and vary only the random participation masks, comparing zero-filled and observed-only representations. As shown in Table 2, at 20% participation, the constant-signal frequency score changes from 0.1500 (observed-only) to 0.2327 (constant zero-filled), whereas both approaches yield 0.1500 at full participation. This confirms and quantifies the influence of participation encoding on frequency scores.

Table 2. Synthetic participation-mask controls.

ParticipationEligible masks (%)Constant Zero-filledConstant Observed-onlySinusoid Zero-filledSinusoid Observed-only
10%38.36 ± 0.800.2307 ± 0.00450.1500 ± 0.00000.2294 ± 0.00430.4868 ± 0.0112
20%92.88 ± 0.900.2327 ± 0.00080.1500 ± 0.00000.2314 ± 0.00100.3707 ± 0.0099
40%100.00 ± 0.000.2364 ± 0.00250.1500 ± 0.00000.2347 ± 0.00220.2443 ± 0.0028
100%100.00 ± 0.000.1500 ± 0.00000.1500 ± 0.00000.5000 ± 0.00000.5000 ± 0.0000
AuthorsOpenReview ↗

Response by Authors (Part 2)

W2 & Q2 (cont.)

End-to-end validation. To determine whether these spectral effects undermine defense effectiveness, we provide clients’ weights directly in Table 3. Under both matched alternating attacks, zero filling retains more benign weight, assigns less malicious weight, lowers FPR and raises TPR relative to observed-only encoding. These results support the timeline-preserving representation in these controls.

Table 3. Aggregation and detection diagnostics under alternating attacks.

AttackTrajectoryBenign weight ↑Malicious weight ↓FPR (%) ↓TPR (%) ↑
Hist-ResidualZero-filled0.8893 ± 0.00370.4548 ± 0.01270.207 ± 0.1243.236 ± 0.988
Hist-ResidualObserved-only0.8475 ± 0.00820.6913 ± 0.00780.439 ± 0.1300.694 ± 0.293
Profile-LIEZero-filled0.8906 ± 0.00430.4567 ± 0.01320.172 ± 0.0493.444 ± 1.076
Profile-LIEObserved-only0.8498 ± 0.00760.6936 ± 0.00660.467 ± 0.1270.840 ± 0.177

Q3: Compact Representation and Information Loss

Our compact representation is designed to retain essential update characteristics while substantially reducing the computational overhead of risk estimation. FedRASR preserves all 130 forecasting-head coordinates and summarizes the remaining 23,968 backbone coordinates using four pooled means and one RMS magnitude per tensor, resulting in a 190-dimensional representation. This design retains complete forecasting-head information and captures the statistical characteristics and magnitudes of all backbone tensors. The full update's squared norm remains recoverable from the retained coordinates and RMS statistics.

The dimensionality reduction does not imply an equivalent loss of discriminative information. The compression primarily discards fine-grained coordinate patterns within pooling groups, while preserving the aforementioned statistical information. Moreover, the compact representation is used only for risk estimation, whereas the complete model updates are retained for final aggregation.

Regarding potential evasion, theoretically, we acknowledge that perturbations preserving the forecasting-head coordinates, pooled means, and per-tensor RMS magnitudes would produce identical compact representations and thus be indistinguishable at the descriptor level. However, in practice, this represents a stronger, specifically tailored adaptive attack beyond our evaluated poisoning scenarios. We will discuss this corner-case attack and provide this limitation in the revised paper.

W3 & Q4: Regional Calibration and Historical Feedback

Our threat model is based on a fixed malicious-client pool randomly sampled from the global population, with a 30% malicious-client ratio (to satisfy the benign-majority assumption), following the commonly adopted setting in robust federated learning [1, 2]. Furthermore, only 20% of all clients participate in each communication round. Under random malicious-client allocation and random participation, the probability that malicious clients dominate a particular region's participating clients is very low. Therefore, deliberate regional adversarial dominance is unlikely under our considered threat model and represents a more extreme adversarial setting.

Nevertheless, to directly address this concern, we further consider a challenging scenario where 49% of clients in a specific region (Portugal) are malicious (Must < 50% to follow the protocol). This setting substantially increases regional adversarial concentration compared with our original random-compromise setting. In the tables below, FedRASR still works and achieves the lowest mean RMSE and MAE among the three compared methods and these results demonstrate FedRASR's robustness under concentrated regional poisoning.

Hist-Residual

RegionRMSE FedRASR ↓RMSE FedAvg ↓RMSE FreqFed ↓MAE FedRASR ↓MAE FedAvg ↓MAE FreqFed ↓
Portugal0.129340.159130.137010.112570.138720.12754
Ausgrid0.083550.084780.084250.061500.074130.06378
London0.087750.092830.090930.062380.076790.07254
Swiss0.103200.107620.104280.070320.090370.08461
Four-region macro0.100960.111090.104120.076690.095000.08712

Profile-LIE

RegionRMSE FedRASR ↓RMSE FedAvg ↓RMSE FreqFed ↓MAE FedRASR ↓MAE FedAvg ↓MAE FreqFed ↓
Portugal0.132740.160620.147290.110310.141830.12015
Ausgrid0.081760.084300.083930.062140.073280.06851
London0.088020.090410.089650.063920.078970.07253
Swiss0.096510.104130.098730.072160.084460.07882
Four-region macro0.099760.109870.104900.077130.094640.08500
AuthorsOpenReview ↗

Response by Authors (Part 3)

W3 & Q4 (cont.)

Regarding benign distribution shifts, FedRASR does not rely on fixed regional references. Instead, both regional centers and risk thresholds are dynamically updated using exponential moving averages (EMA). This design balances stability against short-term fluctuations with gradual adaptation to evolving regional data distributions, reducing the sensitivity of risk estimation to transient benign changes.

Historical soft weights are used only for selecting the risk-threshold reference set. Regional centers are estimated using coordinate-wise medians of all currently participating clients within each region. For threshold calibration, according to Appendix E.4, clients with historical soft weights ≥ 0.85 are prioritized. If fewer than 12 eligible clients are available, the reference set falls back to all current regional participants. Together with EMA smoothing, these mechanisms help maintain reference availability and reduce sensitivity to individual historical decisions.

[1] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. Dba: Distributed backdoor attacks against federated learning. In International conference on learning representations. 2019

[2] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International conference on machine learning. 2018

W4 & Q5: Soft Reweighting and Adversarial Influence Bound

As clarified in W1/Q1, FedRASR aims to continuously adjust client contributions based on estimated risks, rather than making permanent binary decisions about benign and malicious clients. From an information-preservation perspective, hard isolation has lower Shannon entropy and completely discards flagged updates, potentially losing valuable information from misclassified benign clients. In contrast, soft reweighting has higher Shannon entropy and preserves graded contributions from a broader set of clients, while continuously penalizing suspicious updates, including those that do not trigger binary detection. The positive weight floor is therefore a deliberate design choice to balance information preservation and malicious influence suppression. Regarding the theoretical influence bound, we provide a brief proof. Let pi=niai/∑jnjajp_i=n_i a_i/\sum_j n_j a_j denote the normalized aggregation weight and ηM=∑i∈Mpi\eta_{\mathcal M}=\sum_{i\in\mathcal M}p_i denote the total malicious aggregation share. Under the condition that malicious update norms are bounded by ∣Δi∣≤B|\Delta_i|\leq B, Eq. (15) yields: $ \left|\sum_{i\in\mathcal M}p_i\Delta_i\right| \leq B\eta_{\mathcal M}. $ This establishes a conditional upper bound on the direct malicious aggregation contribution, showing that reducing the malicious aggregation share tightens its influence bound. As stated in W1/Q1, experimental results in Appendix Table S9 validate the practical benefits of soft reweighting in preserving benign contributions while more effectively suppressing malicious influence under the evaluated attacks.

W5 & Q6: Component Design and Extended Evaluation

FedRASR is motivated by three key design objectives: (i) mitigating benign regional heterogeneity through regional calibration, (ii) capturing persistent malicious behaviors through cross-round frequency-domain analysis, and (iii) preserving useful client information while suppressing suspicious updates through risk-aware soft reweighting. These components are designed to work collaboratively, instead of relying on individual heuristics or binary malicious-client classification. We have actually conducted the ablation study in Appendix Table S9, comparing full FedRASR with Hard Isolation, Time-domain Scoring, and Basic Frame variants. FedRASR consistently outperforms ablations under both attacks. Comparison with Hard Isolation supports the benefits of soft reweighting, while comparison with Time-domain Scoring supports the effectiveness of frequency-based design.

We use a single frozen configuration for the frequency-score weights, trajectory window, threshold-calibration parameters, rank gates, and soft-weight floor across all four regions. Although regional centers and numerical thresholds are estimated separately from each region's observed distributions, they follow the same calibration rules. This demonstrates that FedRASR can operate across heterogeneous regions without region-specific manual hyperparameter tuning.

Reviewer_V7e2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FedRASR defends federated load forecasting by tracking compact client-update features, subtracting regional reference trajectories, and converting frequency and directional evidence into soft aggregation weights. On four regional smart-meter collections, it improves forecasting error under two persistent poisoning attacks while staying close to FedAvg on clean data.

Critical strengths

Regional calibration addresses a concrete difficulty: benign differences between regions can resemble anomalous updates. The reported gains come with little clean-performance loss, and the five-seed experiments and ablations support contributions from both frequency scoring and soft reweighting within the evaluated setting.

Critical weaknesses
  1. Both attacks share the odd/even-round sign schedule of Hist-Residual (Eq. 4); Profile-LIE adds update shaping to the same corruption. This creates the alternating behavior that the frequency and directional scores are designed to detect. Without constant-sign, slowly varying, randomized, or defense-aware attacks, it is hard to judge robustness beyond this schedule. Zero-filling roughly 80% of trajectory positions also mixes participation patterns with update dynamics, with no control separating the two.

  2. Appendix C.2 says the reported evaluation uses the validation segment, and Table S2 limits it to ten clients per region. Results on the untouched test split with broader client coverage are needed to assess generalization.

  3. The supplied FoolsGold code omits the original logit reweighting, while WAFFLE-FT replaces WAFFLE's pre-training data-representation detector with per-round FFT-energy filtering of updates. These adaptations need validation. FLDetector, FreqFed and FedRDF are cited but not evaluated. Recent related defenses, including AdaAggRL (AAAI 2025) and FLGMM (Information Fusion 2026), are also missing

Questions for authors
  1. Does the gain survive removal of the alternating sign schedule and a mask-aware or observed-update-only trajectory control, with attack strength and participation held fixed?

  2. Can you evaluate frozen hyperparameters on untouched test segments with broader client coverage, and clarify how validation data were used for tuning?

  3. Can you validate the FoolsGold/WAFFLE adaptations and compare faithful, compatible recent and history/frequency-based defenses under the same protocol?

3 comments
AuthorsOpenReview ↗

Response by Authors (Part 1)

We are pleased that the reviewer acknowledges the significance of regional calibration, the competitive performance under clean training, and the evidence from five random seeds. Related weaknesses and questions are addressed together below.

W1 & Q1: Attack Generalization and Participation-Mask Effects

We thank the reviewer for highlighting two important concerns: (i) whether FedRASR's effectiveness depends on the alternating sign schedule shared by both attacks, and (ii) whether zero-filled missing rounds introduce artificial frequency patterns that affect risk estimation.

We agree that evaluating only alternating attacks limits the generalizability of FedRASR. Therefore, we additionally evaluate constant-sign and randomized-sign attacks, and compare zero-filled and observed-only trajectories. As shown in Table 1, zero-filled FedRASR achieves the lowest mean RMSE and MAE among the three compared methods across all six settings, demonstrating that its advantage is not restricted to alternating attacks.

Table 1. Attack schedule and trajectory comparisons.

AttackSignFedAvg RMSE / MAE ↓FedRASR Zero-filled RMSE / MAE ↓FedRASR Observed-only RMSE / MAE ↓
Hist-ResidualAlternating0.1037 ± 0.0024 / 0.0807 ± 0.00190.0937 ± 0.0018 / 0.0683 ± 0.00220.1009 ± 0.0017 / 0.0776 ± 0.0010
Hist-ResidualConstant0.0829 ± 0.0015 / 0.0517 ± 0.00100.0828 ± 0.0018 / 0.0514 ± 0.00130.0831 ± 0.0014 / 0.0515 ± 0.0009
Hist-ResidualRandomized0.0947 ± 0.0104 / 0.0690 ± 0.01460.0945 ± 0.0083 / 0.0689 ± 0.01140.0948 ± 0.0092 / 0.0695 ± 0.0128
Profile-LIEAlternating0.1031 ± 0.0022 / 0.0787 ± 0.00170.0940 ± 0.0021 / 0.0683 ± 0.00210.1008 ± 0.0017 / 0.0764 ± 0.0013
Profile-LIEConstant0.0837 ± 0.0018 / 0.0525 ± 0.00140.0834 ± 0.0018 / 0.0520 ± 0.00150.0839 ± 0.0016 / 0.0522 ± 0.0012
Profile-LIERandomized0.0948 ± 0.0107 / 0.0686 ± 0.01430.0945 ± 0.0084 / 0.0685 ± 0.01110.0946 ± 0.0092 / 0.0686 ± 0.0124

Although Table 1 favors zero-filled trajectories, it does not establish whether zero filling introduces artificial frequency patterns. We therefore conduct synthetic experiments using constant benign residuals and sinusoidal residuals under different participation rates. As shown in Table 2, at 20% participation, zero filling increases the constant-signal frequency score from 0.1500 to 0.2327, confirming that missing-round encoding affects spectral characteristics.

Table 2. Synthetic participation-mask controls.

ParticipationEligible masks (%)Constant Zero-filledConstant Observed-onlySinusoid Zero-filledSinusoid Observed-only
10%38.36 ± 0.800.2307 ± 0.00450.1500 ± 0.00000.2294 ± 0.00430.4868 ± 0.0112
20%92.88 ± 0.900.2327 ± 0.00080.1500 ± 0.00000.2314 ± 0.00100.3707 ± 0.0099
40%100.00 ± 0.000.2364 ± 0.00250.1500 ± 0.00000.2347 ± 0.00220.2443 ± 0.0028
100%100.00 ± 0.000.1500 ± 0.00000.1500 ± 0.00000.5000 ± 0.00000.5000 ± 0.0000

To determine whether these spectral effects undermine defense effectiveness, we provide clients’ weights directly. Under both alternating attacks, zero filling retains more benign weight, assigns less malicious weight, lowers FPR and raises TPR relative to observed-only encoding. These results support the timeline-preserving representation in these controls.

Table 3. Aggregation and detection diagnostics under alternating attacks.

AttackTrajectoryBenign weight ↑Malicious weight ↓FPR (%) ↓TPR (%) ↑
Hist-ResidualZero-filled0.8893 ± 0.00370.4548 ± 0.01270.207 ± 0.1243.236 ± 0.988
Hist-ResidualObserved-only0.8475 ± 0.00820.6913 ± 0.00780.439 ± 0.1300.694 ± 0.293
Profile-LIEZero-filled0.8906 ± 0.00430.4567 ± 0.01320.172 ± 0.0493.444 ± 1.076
Profile-LIEObserved-only0.8498 ± 0.00760.6936 ± 0.00660.467 ± 0.1270.840 ± 0.177
AuthorsOpenReview ↗

Response by Authors (Part 2)

W2 & Q2: Evaluation Protocol and Generalization

In the experiment, the dataset is chronologically divided into 80% training, 10% validation, and 10% testing. The training set is used for model optimization, the validation set for hyperparameter tuning, and the test set for final performance evaluation, which is a widely used mechanism in ML training. We acknowledge that the wording in Appendix C.2 is inappropriate and may have caused confusion. Actually, the validation set is not used for model training or final testing, but only for tuning hyperparameters, including the trajectory window, sketch dimension, and soft-weighting parameters. The tuning process considers forecasting RMSE, benign FPR, benign weight preservation, and malicious aggregation mass.

To further assess generalization, we expand the evaluation from 10 to 300 clients per region. As shown below, FedRASR outperforms baselines, while maintaining comparable clean performance.

Table 4. Results of 300 clients per region.

MethodRMSE Clean ↓RMSE Hist-Residual ↓RMSE Profile-LIE ↓MAE Clean ↓MAE Hist-Residual ↓MAE Profile-LIE ↓
FedRASR (ours)0.0897 ± 0.00220.0955 ± 0.00220.0950 ± 0.00200.0628 ± 0.00170.0719 ± 0.00290.0703 ± 0.0024
FedAvg0.0897 ± 0.00230.1090 ± 0.00190.1064 ± 0.00200.0628 ± 0.00160.0890 ± 0.00170.0846 ± 0.0017
Krum0.0954 ± 0.00480.1115 ± 0.00870.1126 ± 0.01610.0647 ± 0.00460.0784 ± 0.00730.0798 ± 0.0144
Trimmed Mean0.0909 ± 0.00190.1050 ± 0.00380.1068 ± 0.00480.0640 ± 0.00130.0796 ± 0.00400.0812 ± 0.0046
GeoMed0.0904 ± 0.00240.1049 ± 0.00290.1020 ± 0.00210.0631 ± 0.00200.0826 ± 0.00310.0782 ± 0.0021
FoolsGold-style0.0914 ± 0.00110.1074 ± 0.00070.1063 ± 0.00120.0643 ± 0.00110.0866 ± 0.00080.0839 ± 0.0007
WAFFLE-FT0.0899 ± 0.00250.1046 ± 0.00740.1063 ± 0.00500.0630 ± 0.00200.0848 ± 0.00940.0859 ± 0.0059
IF-Detector0.0899 ± 0.00250.1118 ± 0.00590.1084 ± 0.00630.0630 ± 0.00200.0871 ± 0.00620.0823 ± 0.0062

W3 & Q3: Baseline Fairness and Extended Comparisons

We thank the reviewer for pointing out the baseline implementation and coverage. In our previous implementations of FoolsGold and WAFFLE, we retained their core algorithmic components while omitting certain modules that were not directly applicable to our federated load forecasting task or could introduce inconsistencies in the experimental protocol. As requested, we have reimplemented FoolsGold and WAFFLE following their original methodologies.

Furthermore, we introduce five additional baselines, FreqFed, FLGMM, FedRDF, FLDetector and AdaAggRL.

The results with the currently reported baselines, and reimplemented FoolsGold and WAFFLE are shown in the table below.

Table 5. Extended baseline comparisons.

MethodRMSE Clean ↓RMSE Hist-Residual ↓RMSE Profile-LIE ↓MAE Clean ↓MAE Hist-Residual ↓MAE Profile-LIE ↓
FedRASR (ours)0.0897 ± 0.00220.0955 ± 0.00220.0950 ± 0.00200.0628 ± 0.00170.0719 ± 0.00290.0703 ± 0.0024
FoolsGold (corrected)0.0956 ± 0.00250.1087 ± 0.00330.1089 ± 0.00330.0679 ± 0.00250.0856 ± 0.00300.0846 ± 0.0027
FreqFed0.0989 ± 0.00780.1044 ± 0.00990.0956 ± 0.00430.0696 ± 0.00630.0769 ± 0.01190.0760 ± 0.0038
FLGMM0.0918 ± 0.00300.1087 ± 0.00380.1071 ± 0.00620.0641 ± 0.00200.0848 ± 0.00450.0832 ± 0.0089
WAFFLE-FT (official)0.0898 ± 0.00310.1103 ± 0.00270.1073 ± 0.00240.0629 ± 0.00210.0882 ± 0.00300.0840 ± 0.0020
FLDetector0.0928 ± 0.00290.1039 ± 0.00320.1063 ± 0.00550.0646 ± 0.00260.0779 ± 0.00320.0798 ± 0.0052
FedRDF0.0931 ± 0.00290.1036 ± 0.00350.1056 ± 0.00470.0651 ± 0.00230.0776 ± 0.00340.0791 ± 0.0046
AdaAggRL0.0950 ± 0.00350.1024 ± 0.00410.0994 ± 0.00360.0709 ± 0.00380.0788 ± 0.00400.0837 ± 0.0032

Dear Reviewer:

Thank you very much for your constructive comments. We hope that our responses have addressed your concerns. Please feel free to let us know if any further clarification is needed.

Best regards,

The authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary and General Response to the Reviewers

We sincerely thank all reviewers for their time and constructive feedback and for recognizing the importance of robust federated load forecasting. Our rebuttal clarifies the motivation and operating assumptions of FedRASR, the compact update representation, and the training, validation, and testing protocol. We add constant-sign and randomized-sign attacks, and experiments with concentrated regional contamination. Furthermore, we expand held-out evaluation to 300 clients per region, and also add five new baseline comparisons, plus additional FoolsGold and WAFFLE reimplementations. The experimental results support FedRASR's robustness under the evaluated attacks while maintaining competitive clean performance. The revisions and additional experiments have been incorporated into the updated paper, highlighted in blue color.

We would greatly appreciate the reviewers' feedback on whether our responses and additional experiments address their main concerns. We would also be happy to clarify any remaining questions.

AuthorsOpenReview ↗

Summary of the Rebuttal by Authors

We thank all reviewers for their constructive comments and summarize our responses to each concern below.

Reviewer V7e2

W1 & Q1: Alternating attacks and mask effects

A: Added constant-sign and randomized-sign attacks, plus participation-mask controls in Tables S15–S17.

W2 & Q2: Expanding evaluation results

A: Clarified the data split and expanded testing to 300 clients per region in Tables S13 and S14.

W3 & Q3: Baseline fairness and coverage

A: Reimplemented FoolsGold and WAFFLE and added FreqFed, FLGMM, FedRDF, and FLDetector in Table S14.

Reviewer 4Hfm

W1 & Q1: Risk score rationale and motivation

A: Clarified soft reweighting and its motivation, with ablations in Appendix Table S9.

W2 & Q2: Zero-filling and mask effects

A: Added synthetic mask controls and zero-filled versus observed-only diagnostics (Tables S16 and S17).

Q3: Information loss from compression

A: Clarified what the 190-dimensional representation retains and its limitations against adaptive attacks.

W3 & Q4: Regional contamination and historical feedback

A: Tested 49% malicious clients in Portugal and clarified reference updates and fallback.

W4 & Q5: Soft weights and influence bound

A: Explained the weight floor and provided a conditional bound on malicious influence.

W5 & Q6: Component design and extended evaluation

A: Clarified component roles and added broader evaluations.

Reviewer Vgzo

W1: Dependence on alternating attacks

A: Added constant-sign and randomized-sign attacks.

W2: Concentrated regional contamination

A: Added experiments with 49% malicious clients in Portugal.

W3: Performance differences across regions

A: Confirmed RMSE improvements across all four regions on the expanded test cohort.

W4: Limited baseline comparisons

A: Added FreqFed, FedRDF, and FLDetector comparisons.

Dear reviewers, as the deadline is approaching, please check the rebuttal and consider participating in the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.