acceptodds
Withdrawn from ICLR 2027

Federated Spatiotemporally Regularized Latent Factorization of Tensors for Privacy-Preserving Multi-Indicator Data Recovery in Wireless Sensor Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Wireless sensor networks (WSNs) continuously collect multivariate environmental data, yet sensor failures, energy constraints, and unreliable communication frequently lead to substantial missing values. Latent factor analysis (LFA) is a representative approach to extract compact latent representations from irregular environmental sensing data, providing informative features for downstream data recovery. However, existing LFA methods typically model each sensing indicator independently, overlooking the inherent spatiotemporal correlations across indicators in sensing data. Moreover, they require raw data to be aggregated at a central server, conflicting with increasingly stringent privacy-preserving requirements. To solve these issues, this paper proposes a federated spatiotemporally-regularized latent factorization of tensors (FedSR-LFT) model. It builds on two-fold key ideas: 1) developing a sensor-level federated latent factorization of tensors model that jointly models multiple sensing indicators and supports privacy-preserving collaborative training by exchanging only gradient information while keeping raw data local, and 2) incorporating spatiotemporal regularization and learnable indicator-loss weight to enrich latent spatiotemporal information and adaptively adjust each indicator's influence, thereby enabling accurate multi-indicator data recovery. Together, these designs enable FedSR-LFT to exploit cross-indicator dependencies for accurate WSN data recovery with privacy preservation. Extensive experiments on two real-world WSN datasets are conducted to compare FedSR-LFT with eight state-of-the-art federated signal recovery baselines. The results show that FedSR-LFT significantly outperforms all baselines in recovery accuracy while preserving data privacy. The source code and datasets are submitted as Supplementary Materials for Reproducibility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Privacy-Preserving, Wireless Sensor Networks, Latent Factorization of Tensor, Spatiotemporal Correlation, Data Recovery

Reviews

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

This paper proposes FedSR-LFT, a federated tensor factorization framework for missing data recovery in wireless sensor networks. It combines multi-indicator modeling, spatial and temporal regularization, and adaptive indicator weights. Experiments on two real-world datasets report improved recovery accuracy and resistance to gradient inversion attacks.

Critical strengths
  1. Effective integration of spatiotemporal information. Jointly modeling multiple sensing indicators with spatial and temporal regularization is well motivated. The reported improvements over existing federated baselines are substantial, particularly under high missing rates.
  2. Reasonable experimental coverage. The paper includes comparisons with eight baselines, parameter sensitivity studies, and gradient inversion experiments. These provide useful evidence of the proposed framework's potential.
Critical weaknesses
  1. Questionable privacy guarantees. The framework requires clients to obtain neighboring sensors' latent vectors for spatial regularization, while simultaneously claiming that these representations remain private. This information-sharing mechanism is not sufficiently explained. Moreover, the privacy analysis mainly argues that exact reconstruction is underdetermined, which does not establish a rigorous privacy guarantee. The DLG experiments alone are insufficient to support the broader privacy claims.
  2. Limited novelty and experimental generalization. The proposed components largely build upon established tensor factorization, graph regularization, and federated optimization techniques. The experiments cover only two datasets with randomly simulated missing values. More importantly, the contribution of adaptive indicator weighting is not independently demonstrated through a dedicated ablation. These limitations make it difficult to establish the broader significance of the proposed approach.
Questions for authors

Q1. How are neighboring sensors' latent vectors exchanged without violating the claimed privacy constraints? Can the authors clarify the threat model and provide stronger privacy guarantees beyond the DLG evaluation?

Q2. Can the authors provide a dedicated ablation of adaptive indicator weighting and evaluate the method under structured missingness, such as consecutive sensor failures? These results would help establish the individual contributions and practical robustness of the proposed framework.

1 comment

We sincerely thank the reviewer for the constructive feedback. We have addressed each weakness (W) and question (Q) through the responses and improvements detailed below. The corresponding discussions will also be incorporated into the revised manuscript.

W1.1 & Q1.1 Regional communication mechanism. FedSR-LFT assumes mutual trust among sensors within each region. Each sensor obtains the latent factors needed for spatial regularization from other sensors in its region. These factors are shared and retained within the region, without being uploaded to the central server. Raw sensing data remain at their respective sensors. The server sends the shared matrices VV and WW and indicator weights to sensors, and receives V,WV,W gradients and indicator-wise aggregated absolute errors. Thus, sensor latent factors remain private with respect to the central server while being available for spatial regularization within the trusted region.

W1.2 & Q1.2 Clarification of the threat model In the DLG-based attack, the attacker initializes dummy sensing data and sensor factors, computes dummy time and indicator gradients, and jointly minimizes their discrepancies from the uploaded gradients. It also matches the dummy data's total absolute error to the sum of uploaded total absolute errors. We agree that an underdetermined reconstruction problem does not establish a rigorous privacy guarantee, and our wording overstated this conclusion. We will revise the privacy claims to reflect the evidence from the evaluated attacks and clarify the threat model and attack settings. The appendix also evaluates clipping and Gaussian perturbation as additional protection: each time-gradient row is clipped to an L2 norm of at most CC, then perturbed with independent Gaussian noise of standard deviation nm×C\mathrm{nm}\times C. Under the reported DLG setting, this increases attack reconstruction RMSE and MAE while maintaining competitive recovery accuracy. These results demonstrate an empirical privacy–accuracy trade-off.

W2.1 & Q2.1 Technical contributions and learnable indicator-loss weights. We appreciate the opportunity to clarify the technical contributions of FedSR-LFT, which lie in two key ideas: (1) FedSR-LFT introduces indicator factors to model a sensor–time–indicator tensor within a sensor-level federated framework, explicitly capturing cross-indicator dependencies, and accordingly develops learnable indicator-loss weights to balance indicator contributions during collaborative recovery. (2) We innovatively propose a regional spatiotemporal regularization strategy specifically designed for multi-indicator WSN data recovery under the proposed sensor-level federated tensor framework.

Furthermore, we have conducted an ablation experiment to assess the contribution of the learnable indicator-loss weights. We fixed the indicator-loss weights at${1,1,1}$ while keeping all other experimental settings unchanged. The results below report the overall RMSE and MAE under the two weight settings.

Datasetρ\rhoLearnable weightsUniform weights
Beijing Air Quality0.00156.511/28.24957.120/28.772
0.0139.876/19.08340.496/19.615
Intel Lab0.0014.213/2.2874.289/2.328
0.012.480/0.7752.485/0.801

The learnable indicator-loss weights achieve lower overall RMSE and MAE in all four settings. These consistent improvements support the contribution of dynamically adjusting indicator-loss weights to recovery accuracy.

W2.2 & Q2.2 Evaluate the method under structured missingness. Starting from both datasets at$\rho=0.01$, we constructed two missingness patterns: (1) Contiguous temporal gap: removing training data within a contiguous window covering 50% of the time slots for each of 20% of sensor–indicator sequences; and (2) Sensor outage: removing training data for all indicators within a common contiguous window covering 50% of the time slots for each of 20% of sensors. The results below show that FedSR-LFT achieves lower RMSE and MAE than FLFL in all 24 comparisons, demonstrating that its performance advantage persists under both structured missingness patterns.

PatternDatasetIndicatorFLFLFedSR-LFT
Temporal gapBeijing Air QualityPM10112.681/74.80660.724/36.210
PM2.579.871/48.98144.005/27.511
CO1.413/0.8420.874/0.491
Intel LabTemperature14.416/11.3933.961/1.165
Humidity13.702/10.4442.399/1.551
Voltage0.821/0.5750.042/0.028
Sensor outageBeijing Air QualityPM10110.439/72.87959.797/35.779
PM2.580.861/49.61343.604/27.107
CO1.420/0.8480.869/0.485
Intel LabTemperature14.408/11.3713.921/1.070
Humidity13.870/10.5722.076/1.360
Voltage0.804/0.5670.029/0.020
Reviewer_3sTYOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedSR-LFT, a federated tensor factorization method for recovering missing measurements in wireless sensor networks. It jointly models multiple measurement types and combines spatial and temporal regularization with adaptive indicator weights. The method aims to keep raw measurements local while coordinating training through exchanged updates. Validation uses two environmental sensing datasets, eight federated baselines, component ablations, a comparison with centralized training, and gradient-reconstruction attack experiments. The paper reports consistently lower recovery errors and improved resistance to the evaluated attacks.

Critical strengths
  1. The approach is well motivated for missing sensor measurements, especially since related indicators, nearby sensors, and adjacent time points can all provide useful information for recovering missing values. Combining these relationships in a joint model is sensible, and the ablations in Figures 4 and 7 show that both spatial and temporal regularization reduce recovery error in the tested settings.
  2. The reported recovery improvements are substantial and consistent throughout the paper and FedSR-LFT achieves the lowest reported RMSE and MAE across the evaluated datasets and sampling rates (Tables 1, 6 and 7). At a 10% sampling rate, PM2.5 RMSE decreases from 45.497 for FLFL to 31.964, while Intel Lab humidity RMSE decreases from 1.957 to 0.914. The federated and centralized versions also produce very similar mean errors in the comparison reported in Table 8.
Critical weaknesses
  1. Section 3.5 and Algorithm 1 require sensors to obtain neighboring sensors’ latent factors, while Appendix B.4 assumes these factors are unavailable to the server. How this exchange works is therefore important but insufficiently explained. Although the analysis discusses consecutive training rounds, the fact that gradients depend on unknown factors does not by itself establish that those factors or the underlying measurements cannot be inferred. The DLG experiments provide useful evidence about the tested attack, but I am not convinced they support the broader privacy claims as currently stated.
  2. The experiments demonstrate the benefits of spatial and temporal regularization, but these are already present in FLFL. Separate comparisons are needed to establish how much joint indicator modeling and adaptive weighting contribute. This is particularly relevant because the indicators have different units and numerical ranges, while the weighting rule depends on their absolute reconstruction errors.
Questions for authors
  1. How do sensors obtain neighboring sensors’ latent factors without exposing them to the central server? Please clarify the communication mechanism and whether the attack experiments use all information available to the server across training rounds.
  2. Could you compare adaptive versus uniform indicator weights, and joint tensor factorization versus separate single-indicator models with otherwise matched settings? Were indicators standardized before training, and how does this choice affect the results?
  3. Do the improvements over FLFL persist under structured missingness, such as contiguous temporal gaps or outages that remove all indicators from a sensor simultaneously?
2 comments

We sincerely thank the reviewer for the constructive feedback. We have addressed each weakness (W) and question (Q) through the responses and improvements detailed below. The corresponding discussions will also be incorporated into the revised manuscript.

W1 & Q1. Regional communication mechanism and privacy analysis. FedSR-LFT assumes that sensors within each region are mutually trusted. Each sensor obtains the latent factors required from other sensors in the same region to compute spatial regularization. These factors are shared and retained locally within the region and are not uploaded to the central server. Raw sensing data remain at the corresponding sensors. The server sends the shared matrices VV and$W$ and the indicator weights to the sensors, and receives the gradients of VV and$W$ together with the indicator-wise aggregated absolute errors. Therefore, the private sensor latent factors are private with respect to the central server, while being available for spatial regularization within the trusted region.

We consider an honest-but-curious server that attempts to infer sensing data from the information exchanged during training. The attack experiments use time and indicator gradients together with the total aggregated absolute error from a single training round. We additionally evaluated a joint attack using uploads from three consecutive training rounds. The overall RMSE/MAE were 61.235/34.906 on Beijing Air Quality and 22.911/18.196 on Intel Lab. Under this attack configuration, combining three rounds did not improve reconstruction over the corresponding single-round attack. We agree that unknown sensor factors alone do not establish non-reconstructability, and our original privacy claims were too broad. We will revise these statements to reflect the evidence provided by the analysis and attack experiments, and clarify the threat model and attack settings in the revised manuscript.

W2.1 & Q2.1 Contributions of multi-indicator modeling and learnable indicator-loss weights. We sincerely thank the reviewer for pointing out this deficiency. Following this suggestion, we conducted two controlled comparisons: (1) FedSR-LFT versus separate single-indicator models; and (2) learnable indicator-loss weights versus fixed uniform weights.

For experiment (1), each single-indicator model uses only the data of its target indicator, whereas FedSR-LFT jointly models all three indicators. The results below show that FedSR-LFT achieves lower MAE and RMSE in all 24 indicator–metric comparisons. At$\rho=0.001$, it reduces MAE by 29.6–91.6% and RMSE by 30.1–87.8%. The gains are particularly pronounced under severe data sparsity, indicating that correlations among indicators provide complementary information that is unavailable when each indicator is modeled independently.

Datasetρ\rhoIndicatorSingle indicatorFedSR-LFT
Beijing Air Quality0.001PM1086.695/129.22749.816/82.574
PM2.561.799/101.42241.814/63.304
CO0.910/1.6530.640/1.155
Beijing Air Quality0.01PM1040.034/68.57235.694/59.736
PM2.526.833/45.19726.817/43.352
CO0.482/0.9020.479/0.866
Intel Lab0.001Temperature16.798/17.6452.658/4.671
Humidity29.272/30.2303.992/5.513
Voltage2.118/2.1430.178/0.261
Intel Lab0.01Temperature2.880/5.1641.011/3.809
Humidity3.690/4.6711.283/1.964
Voltage0.160/0.2060.020/0.031

For experiment (2), all indicator-loss weights in the uniform variant were fixed at 1 throughout training, consistent with the constraint that their sum equals$K=3$, while all other settings remained unchanged. The following results show that the learnable weights achieve lower overall RMSE and MAE in all eight metric comparisons. Compared with uniform weights, they reduce RMSE by 0.2–1.8% and MAE by 1.8–3.2%. The consistent gains across both datasets and sampling rates demonstrate that dynamically adjusting the indicator-loss weights improves recovery performance over assigning equal importance to all indicators.

Datasetρ\rhoLearnable weightsUniform weights
Beijing Air Quality0.00156.511/28.24957.120/28.772
Beijing Air Quality0.0139.876/19.08340.496/19.615
Intel Lab0.0014.213/2.2874.289/2.328
Intel Lab0.012.480/0.7752.485/0.801

W2.2 & Q2.2 Standardization processing of the indicators before training. Regarding data preprocessing, we applied min-max normalization to each indicator separately, using fixed bounds to map all indicators to the same numerical range. The same preprocessed datasets were used for FedSR-LFT and all baseline models to ensure fair comparison. This normalization makes the indicator-wise absolute errors comparable when updating the learnable indicator-loss weights. To assess the effect of normalization, we additionally trained FedSR-LFT on original-scale data at ρ=0.01\rho=0.01. On Beijing Air Quality, the CO RMSE/MAE increased from 0.866/0.479 to 1.007/0.579 without normalization. On Intel Lab, the overall RMSE/MAE increased from 2.480/0.775 to 2.957/1.427 under the same hyperparameter settings. These results support the use of indicator-wise normalization for joint multi-indicator training, while showing that its effect depends on the dataset and indicator.

Q3. Performance under structured missingness. Starting from both datasets at ρ=0.01\rho=0.01, we constructed two missingness patterns: (1) Contiguous temporal gap: removing training data within a contiguous window covering 50% of the time slots for each of 20% of sensor–indicator sequences; and (2) Sensor outage: removing training data for all indicators within a common contiguous window covering 50% of the time slots for each of 20% of sensors. The results below show that FedSR-LFT achieves lower RMSE and MAE than FLFL in all 24 comparisons, demonstrating that its performance advantage persists under both structured missingness patterns.

PatternDatasetIndicatorFLFLFedSR-LFT
Temporal gapBeijing Air QualityPM10112.681/74.80660.724/36.210
PM2.579.871/48.98144.005/27.511
CO1.413/0.8420.874/0.491
Intel LabTemperature14.416/11.3933.961/1.165
Humidity13.702/10.4442.399/1.551
Voltage0.821/0.5750.042/0.028
Sensor outageBeijing Air QualityPM10110.439/72.87959.797/35.779
PM2.580.861/49.61343.604/27.107
CO1.420/0.8480.869/0.485
Intel LabTemperature14.408/11.3713.921/1.070
Humidity13.870/10.5722.076/1.360
Voltage0.804/0.5670.029/0.020
Reviewer_GdqaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper innovatively formulates multi-indicator WSN data recovery as a sensor-level federated latent factorization of tensor problem, and proposed FedSR-LFT, enabling joint indicator modeling and privacy-preserving collaborative training. Spatiotemporal regularization and adaptive indicator weights are introduced to capture latent dependencies, balance indicator contributions and improve data recovery accuracy. Extensive experiments on two real-world WSN datasets demonstrate its effectiveness.

Critical strengths
  1. The research is practical and has a clear motivation, which combines privacy-preserving federated learning with multi-metric spatiotemporal data recovery in WSNs. It is a highly relevant and practically significant problem.
  2. Modeling these data as a third-order sensor-time-indicator tensor, and the proposed method is designed to align with the characteristics of the task, making the overall design reasonable.
  3. Extensive experiments on two real-world WSN datasets demonstrate the effectiveness of FedSR-LFT.
Critical weaknesses
  1. The server-side aggregation protocol is unclear. In Section 3.5 about server-side optimization, the authors said after receiving the gradient from sensor n at round r, the server immediately updates the shared latent matrices. This implies asynchronous or sequential SGD updates. However, standard federated learning (such as FedAvg) aggregates updates from multiple clients (e.g., averaging gradients) before updating the global model. If it is a sequential update, the model will be severely biased toward the last sensor participating in training. If it is an averaged update, the notation needs to be corrected.
  2. The novelty is limited. The core components such as tensor decomposition, federated framework and so forth are all very mature techniques. The contribution mainly lies in combining them rather than introducing fundamentally new mechanisms.
  3. The authors said each sensor performs local training using its own data and exchanges only gradient updates. However, sensors are typically resource-constrained devices. Requiring each sensor to perform local training and other operations raises doubts about feasibility.
Questions for authors

According to the above weaknesses, the authors should answer the following questions. The answers for these questions in the rebuttal are crucial to the final decision.

  1. The authors should clarify the server-side aggregation protocol.
  2. The technical novelty of the paper needs to be further clarified.
  3. Please analyze the feasibility of performing training and other operations on sensors.
2 comments

We sincerely thank the reviewer for the thoughtful feedback and constructive suggestions. We have addressed each weakness (W) and question (Q) below through clarifications, additional analyses, and supplementary experiments. We will incorporate the following discussion into the revised paper.

W1 & Q1. Server-side update protocol. The original description in Sec. 3.5 did not explain this protocol clearly. FedSR-LFT uses asynchronous server updates, motivated by differences in sensor computation times and communication availability in practical WSN deployments. Each sensor downloads the latest shared matrices VV and WW when beginning local training and uploads its gradients upon completion. The server applies each received update without waiting for the other sensors, following the asynchronous federated optimization principle [1]. Under this asynchronous protocol, server updates follow the arrival of client updates without a predefined sensor ordering or a fixed last participant.

However, extreme heterogeneity in communication delays and update frequencies can produce stale gradients and unequal sensor influence, potentially biasing optimization toward frequently participating sensors. To examine this condition, we conducted a stress test at ρ=0.001\rho=0.001, in which approximately 25% of the sensors delayed their uploads by 100 global progress cycles. As shown in the following table, asynchronous delay increases all 12 indicator-level errors, with increases of 8.0-24.2% on Beijing Air Quality and 5.4-30.3% on Intel Lab. These results show that severe communication heterogeneity can substantially reduce the recovery accuracy of asynchronous training. FedAvg computes a sample-weighted average of the participating sensors' local updates before updating the shared matrices. The results show that FedAvg is capable of partially alleviating performance degradation caused by extreme communication delays and thus provides a viable alternative under severe communication heterogeneity.

DatasetIndicatorAsyncAsync-delayFedAvg-delay
BeijingPM1082.574/49.81699.891/61.85491.013/57.757‾\underline{\text{91.013/57.757}}
PM2.563.304/41.81475.539/46.55974.605/44.194‾\underline{\text{74.605/44.194}}
CO1.155/0.6401.325/0.691‾\underline{\text{0.691}}1.309‾\underline{\text{1.309}}/0.789
Intel LabTemperature4.671/2.6584.924/2.864‾\underline{\text{2.864}}4.823‾\underline{\text{4.823}}/3.269
Humidity5.513/3.9926.103/4.590‾\underline{\text{6.103/4.590}}7.150/5.072
Voltage0.261/0.1780.330/0.2320.312/0.215‾\underline{\text{0.312/0.215}}

References

[1] Xie C, Koyejo O, Gupta I. Asynchronous Federated Optimization. NeurIPS Workshop on Optimization for Machine Learning (OPT). 2020.

W2 & Q2. Technical novelty. We agree that tensor factorization and federated learning are established techniques, and the original manuscript did not sufficiently distinguish these individual ingredients from the technical novelty of FedSR-LFT. We do not claim novelty for these techniques in isolation. Instead, the technical contributions of this work are twofold: (1) FedSR-LFT introduces indicator factors to model a sensor-time-indicator tensor within a sensor-level federated framework for WSN data recovery, explicitly capturing cross-indicator dependencies, and accordingly develops an adaptive indicator weighting mechanism to balance indicator contributions and enable accurate collaborative recovery with privacy preservation. (2) We innovatively propose a regional spatiotemporal regularization straegy specifically designed for multi-indicator WSN data recovery under the proposed sensor-level federated tensor framework.

We have additionally conducted the following two experiments. First, the table below compares learnable indicator-loss weights with fixed uniform weights on both datasets at ρ∈{0.001,0.01}\rho\in\{0.001,0.01\}. In the uniform variant, all indicator weights are fixed at 1 throughout training, while all other experimental settings remain unchanged. The learnable weights achieve lower overall RMSE and MAE in all eight metric comparisons. Relative to uniform weights, they reduce RMSE by 0.2–1.8% and MAE by 1.8–3.2%. These consistent gains across both datasets and sampling rates demonstrate that dynamically adjusting the indicator-loss weights improves recovery performance over assigning equal importance to all indicators.

Datasetρ\rhoAdaptive RMSE/MAEUniform RMSE/MAE
Beijing Air Quality0.00156.511/28.24957.120/28.772
Beijing Air Quality0.0139.876/19.08340.496/19.615
Intel Lab0.0014.213/2.2874.289/2.328
Intel Lab0.012.480/0.7752.485/0.801

Second, to evaluate the contribution of multi-indicator modeling, we compared FedSR-LFT with three single-indicator models on both datasets at$\rho\in{0.001,0.01},whilekeepingallotherhyperparametersettingsunchanged.TheresultsinthetablebelowshowthatFedSR−LFTachieveslowerMAEandRMSEthanthecorrespondingsingle−indicatormodelsinall24comparisons.Theadvantageisparticularlypronouncedat, while keeping all other hyperparameter settings unchanged. The results in the table below show that FedSR-LFT achieves lower MAE and RMSE than the corresponding single-indicator models in all 24 comparisons. The advantage is particularly pronounced at\rho=0.001$, where FedSR-LFT reduces MAE by 29.6–91.6% and RMSE by 30.1–87.8%. This indicates that cross-indicator correlations provide valuable complementary information under severe data sparsity.

Datasetρ\rhoIndicatorSingleFedSR-LFT
Beijing Air Quality0.001PM1086.695/129.22749.816/82.574
PM2.561.799/101.42241.814/63.304
CO0.910/1.6530.640/1.155
Beijing Air Quality0.01PM1040.034/68.57235.694/59.736
PM2.526.833/45.19726.817/43.352
CO0.482/0.9020.479/0.866
Intel Lab0.001Temperature16.798/17.6452.658/4.671
Humidity29.272/30.2303.992/5.513
Voltage2.118/2.1430.178/0.261
Intel Lab0.01Temperature2.880/5.1641.011/3.809
Humidity3.690/4.6711.283/1.964
Voltage0.160/0.2060.020/0.031

W3 & Q3. Feasibility on resource-constrained sensors. Following this suggestion, we analyzed the computation, memory, and communication costs of sensor-side optimization and conducted resource-constrained simulation experiments. FedSR-LFT is designed for sensor nodes that integrate data sensing, local computation, and wireless communication, consistent with recent studies of federated learning on resource-constrained IoT devices [1–3].

For sensor$n$, let$|\Omega_n|$, EE, MM, KK, DD, and$N_g$ denote the number of local data entries, local epochs, time slots, indicators, latent dimension, and sensors in region$g$, respectively. Based on the optimization procedure implemented in FedSR-LFT, the time complexity of one sensor-side update is

O(E∣Ωn∣D+(M+K)D2+NgD).O(E|\Omega_n|D + (M+K)D^2 + N_gD).

The calculation over the local data grows linearly with$|\Omega_n|$ and$D$. The spatiotemporal regularization cost grows linearly with the numbers of time slots and indicators and quadratically with the latent dimension. For example, under the maximum-load Beijing Air Quality setting, ∣Ωn∣=7,047|\Omega_n|=7{,}047, E=1E=1, M=8,736M=8{,}736, K=3K=3, D=20D=20, and$N_g=5$, resulting in an estimated computational scale of approximately 3.64 million operations per sensor-side update.

We further conducted simulation experiments to evaluate the sensor-side computational cost. Sampling rates of 0.01 and 0.3 were selected to represent moderate and maximum local data loads, respectively. For each setting, we evaluated the sensor containing the largest amount of local data. Each sensor performed one local epoch, and its sensor-side update was repeated 30 times under the same single-core environment. The table reports the average execution time over the 30 runs.

DatasetSampling rateLocal data entriesAverage time (ms)Memory (MiB)Upload (MiB)Download (MiB)
Beijing Air Quality0.012607.712.721.3331.334
0.37,0479.313.621.3331.334
Intel Lab0.0155035.406.263.0763.077
0.316,03445.428.263.0773.077

As the sampling rate increases from 0.01 to 0.3, the number of local data entries increases from 260 to 7,047 on Beijing Air Quality and from 550 to 16,034 on Intel Lab. The average sensor-side update time increases only from 7.71 to 9.31 ms and from 35.40 to 45.42 ms, respectively. Under the maximum local data load, the estimated memory requirements remain 3.62 MiB on Beijing Air Quality and 8.26 MiB on Intel Lab.

We also compared these resource requirements with representative commercial sensor-node platforms. The maximum estimated memory requirement of 8.26 MiB is within the 512-MB memory capacity of the Raspberry Pi Zero 2 W, which is equipped with a 1-GHz ARM Cortex-A53 processor. The Beijing Air Quality configurations are also within the memory capacity of the Portenta H7, which provides 8 MB of external SDRAM. These comparisons indicate that the computation and memory requirements of FedSR-LFT can be supported by sensor nodes with local edge-computing capabilities. Sensor nodes with smaller memory capacities can employ single-precision computation, gradient compression, or window-based temporal processing to further reduce the resource requirements.

References

[1] Ficco et al. Federated learning for IoT devices: Enhancing TinyML with on-board training. Information Fusion, 2024.

[2] Deutel et al. On-Device Training of Fully Quantized Deep Neural Networks on Cortex-M Microcontrollers. IEEE TCAD, 2025.

[3] Wong et al. An Empirical Study of Federated Learning on IoT–Edge Devices: Resource Allocation and Heterogeneity. IEEE TNNLS, 2026.

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

Related papers

Discussion (0)

Sign in to comment.