PaperlayerView exact revision on arXiv

Source-linked AI summary

SecFD-FL: Frequency-Decoupled Federated Learning with Secure Aggregation for Charging Load Forecasting in Internet of Electric Vehicles

Mamunur Rashid, Nan Chen, Syed Ali Asad Rizvi, Amiya Nayak

arXiv:2610.12909v1eess.SY

TL;DR

EV charging forecasting faces volatile demand and privacy-sensitive distributed data, while ordinary FL updates remain vulnerable to inversion. SecFD-FL combines frequency-decoupled GRU forecasting with zero-sum secure aggregation; it reports competitive accuracy across real datasets and lower DLG reconstruction correlation.

  • Problem

    High-granularity EV charging loads are volatile and spatially heterogeneous, while shared FL gradients can reveal private samples through inversion attacks.

    Source: §I · p. 1, §II · p. 2

  • Method

    SecFD-FL decomposes charging loads into high- and low-frequency streams for separate GRU processing and applies zero-sum additive masking to client updates.

    Source: §I · p. 2

  • Results

    Across user- and station-level datasets, accuracy is on par with or superior to six FL algorithms; DLG reconstruction correlation is reduced by 8.1 to 15.6 times.

    Source: §I · p. 2, §VI · p. 9

  • Takeaways & Limitations

    The framework reports resistance to gradient inversion without additional per-round communication, beyond a one-time 640-byte key exchange.

    Source: §VI · pp. 9–10

Abstract

from arXiv · show

Accurate charging load forecasting is essential for the reliability and efficiency of the Internet of Electric Vehicles (IoEV), yet it is challenging on two fronts: charging demand is highly stochastic, and the data collected by distributed sensors is privacy-sensitive. Although federated learning (FL) preserves data privacy, collaborative training still remains vulnerable to an honest-but-curious server that can invert shared model updates to reconstruct private data. To address both data volatility and privacy concerns, we propose SecFD-FL, a frequency-decoupled federated learning framework for IoEV comprising: 1) a dual-stream gated recurrent unit (GRU) network that separately models short-term and long-term charging load volatility, and 2) a lightweight secure aggregation layer for privacy preservation. Specifically, the secure aggregation layer employs a zero-sum additive masking protocol that perturbs each client's upload with high-variance noise, preventing the server from reconstructing private data while preserving the exact aggregate. Extensive experiments have been conducted on both EV-level and charging station-level using real datasets. The experiment results show that SecFD-FL reduces global RMSE by up to 60% over centralized baselines, achieving Jain's fairness up to 0.991 across clients, and reduces DLG-attack reconstruction correlation by 8 times to 15 times relative to FedAvg -- all with no additional per-round communication cost beyond a one-time key exchange.

I. INTRODUCTION

EV charging forecasting must contend with volatile, heterogeneous demand while protecting privacy-sensitive data distributed across IoEV clients. SecFD-FL combines frequency-separated forecasting with secure aggregation to address these challenges.

I. INTRODUCTION

  • Charging-demand forecasting is important but difficult because high-granularity loads are volatile and spatially heterogeneous.The introduction situates IoEV sensing across EVs, charging stations, and smart meters as a source of data for real-time operation.Source: §I · p. 1

II. RELATED WORKS

  • Forecasting research progressed from statistical and hand-engineered machine-learning methods toward recurrent deep-learning models, but centralized data aggregation raises privacy concerns.Statistical methods struggle with load nonlinearity and stochasticity, while conventional ML can struggle with long-range temporal dependencies.Source: §I · p. 1, §II · p. 2

I. INTRODUCTION

  • SecFD-FL decomposes charging loads into transient high-frequency and regular low-frequency streams, then learns them independently with a dual-stream GRU.CEEMDAN produces intrinsic mode functions; a zero-sum additive masking protocol protects local updates against gradient inversion.Source: §I · p. 2
  • Global R2 reaches 0.927 on EV WATTS and 0.847 on Palo Alto, and SecFD-FL attains the highest accuracy among six representative federated baselines.The secure-aggregation layer reduces DLG reconstruction correlation by 8.1 to 15.6 times at no additional per-round communication cost.Source: §I · p. 2

II. RELATED WORKS

  • Centralized deep-learning forecasting can model spatio-temporal patterns, but aggregating raw charging records creates privacy and security risks.Prior approaches include convolutional-recurrent, graph-based, feature-fusion, and transformer models.Source: §II · p. 2
  • Existing EV forecasting FL work includes blockchain, decentralized, clustered, transfer-learning, and architectural approaches.The cited studies include Block-FedL, decentralized variants, clustered transfer learning, and station clustering.Source: §II · p. 2
  • Prior FL forecasting methods leave raw-signal volatility, gradient-inversion risk, and limited heterogeneous multi-level benchmarking insufficiently addressed.The paper identifies transient and seasonal patterns as difficult for homogeneous architectures to learn jointly, while shared gradients remain vulnerable to DLG.Source: §I · p. 1, §II · pp. 2–3

A. System Model and Problem Formulation

SecFD-FL trains a shared charging-load forecasting model across clients without sharing raw records. Its system combines local frequency decomposition and dual-stream processing with masked-update aggregation under an honest-but-curious server threat model.

A. System Model and Problem Formulation

  • The federation has K clients that train locally and upload model updates; the server aggregates and broadcasts updates without accessing raw charging records.Clients may represent charging stations or individual EVs, with station-side or user- and fleet-level forecasting use cases.Source: A. System Model and Problem Formulation · p. 3
  • The global model minimizes average next-step charging-load MSE over client data, weighting each client’s local loss by its share of total samples.The resulting objective is the empirical risk over the union of clients’ data.Source: A. System Model and Problem Formulation · p. 3
  • The threat model assumes an honest-but-curious central server that follows aggregation but may infer private data from updates, including through DLG reconstruction.DLG attempts to reconstruct underlying training samples by inverting an observed update.Source: A. System Model and Problem Formulation · p. 3

B. System Architecture of the SecFD-FL Framework

  • The framework processes the decoupled bands through a dual-stream GRU and masks each local update before transmission for secure aggregation.Since the masks sum to zero, the server can perform federated averaging without exposing raw local data.Source: B. System Architecture of the SecFD-FL… · p. 3

C. CEEMDAN-Based Frequency Decoupling

  • Each client applies CEEMDAN to raw load, producing intrinsic mode functions and a residual that reduce mode mixing and yield more stable, separated components.The paper motivates decomposition by the coexistence of slow seasonal trends and sharp transient bursts that a single learner may mix.Source: C. CEEMDAN-Based Frequency Decoupling · p. 3
  • A frequency boundary partitions the padded decomposition into a high-frequency residual band for transient dynamics and a low-frequency trend band for daily or weekly patterns.CEEMDAN’s number of IMFs varies by client; zero-padding gives every client a uniform input shape.Source: C. CEEMDAN-Based Frequency Decoupling · pp. 3–4

B. System Architecture of the SecFD-FL Framework

C. CEEMDAN-Based Frequency Decoupling

D. Dual-Stream GRU Networks

The dual-stream GRU routes distinct frequency bands through independent branches and recombines their predictions into a next-step load forecast.

D. Dual-Stream GRU Networks

  • A parameter-free Lambda layer splits the multi-channel input so each branch processes its designated frequency sub-band independently.The branches use stacked GRUs with intermediate dropout.Source: D. Dual-Stream GRU Networks · p. 4
  • The branches’ forecasted IMF vectors are combined, then inverse-scaled and summed to reconstruct the final load forecast.Fig. 2 shows the outputs concatenated to reconstruct the next-step load.Source: D. Dual-Stream GRU Networks · p. 4
  • The two branches share no weights, so the global parameter vector contains disjoint residual- and trend-stream subsets.This independence allows the two components to be aggregated separately.Source: D. Dual-Stream GRU Networks · p. 4

E. Heterogeneous Federated Aggregation

SecFD-FL uses different aggregation policies for residual and trend streams because they represent statistically distinct phenomena.

E. Heterogeneous Federated Aggregation

1) Stream-specific aggregation weights:

  • The residual stream uses data-proportional client weights because stations with more samples provide statistically stronger transient estimates.Source: 1) Stream-specific aggregation weights · p. 4
  • Trend-stream weights derive from each client’s mean similarity centrality in a graph built from normalized FFT signatures of trend-band mean features.Clients share normalized signatures, which the paper says reveal no time-domain or sample-level information; FFT is used only to derive these weighting signatures, independently of CEEMDAN.Source: 1) Stream-specific aggregation weights · pp. 4–5

2) Stream-specific aggregation rules:

The server aggregates streams separately, retaining part of the previous global residual to limit transient-pattern washout, while zero-sum masks protect individual updates without preventing exact aggregation.

2) Stream-specific aggregation rules:

1) Mask construction:

2) Masked upload and recovery:

Masked uploads preserve exact aggregation, leaving SecFD-FL’s convergence trajectory unchanged; the authors also assess privacy against gradient reconstruction by DLG.

2) Masked upload and recovery:

3) Privacy guarantee against DLG:

  • At SNR ≈2×10^-6, the gradient signal lies five orders below pseudorandom noise, making DLG reconstruction statistically uncorrelated with original charging profiles.Source: 3) Privacy guarantee against DLG · p. 5

G. Overall Algorithm

Algorithm 1 combines one-time signal decomposition with iterative secure federated training, using independent residual and trend updates, stream-specific weights, and zero-sum masking.

G. Overall Algorithm

  • Algorithm 1 integrates one-time signal decomposition with an iterative secure federated training loop and dual-stream updates.Source: G. Overall Algorithm · p. 5
  • Each client applies CEEMDAN, partitions IMFs by a boundary, scales channels, builds τ-step sequences, and sends a spectral signature to the server.Source: G. Overall Algorithm · p. 6
  • After key exchange, training rounds broadcast global weights, train clients locally, generate masks, aggregate stream weights, and update the global model.Source: G. Overall Algorithm · p. 6

A. Datasets

  • The evaluation uses station-level data from Palo Alto and user-level data from EV WATTS.Source: A. Datasets · p. 6

1) Station-level data — City of Palo Alto:

The experiments span station-level and user-level charging data, preprocess session records into hourly sequences, and compare centralized and federated baselines across forecasting utility, fairness, and privacy.

1) Station-level data — City of Palo Alto:

2) User-level data — EV WATTS:

  • EV WATTS records per-vehicle events across a heterogeneous national fleet; 35 eligible vehicles are identified, and five are randomly selected as federated clients.Source: 2) User-level data — EV WATTS · p. 6

B. Data Preprocessing

  • Session energy is distributed across the hours each session spans, then hourly loads are decomposed, scaled, and formed into 24-hour sequences for chronological train/test splits.Source: B. Data Preprocessing · p. 6

C. Baselines

  • SecFD-FL is compared with centralized deep-learning backbones on raw and decomposed inputs and federated optimizers on decomposed signals under a shared training budget.Source: C. Baselines · p. 6

1) Centralized backbones:

2) Federated optimizers:

D. Evaluation Metrics

1) Utility:

  • Utility is measured by MAE, RMSE, and R2 for each client, with federation-level means and standard deviations reported.Source: 1) Utility · p. 7

2) Fairness:

This section defines client fairness and outlines the privacy-attack and communication-cost measurements, alongside the selected frequency boundary.

2) Fairness:

  • Jain’s index over per-client RMSE measures performance equity, with 1 indicating perfectly equal client performance.The privacy evaluation reports gradient-matching loss, reconstruction MSE, and Pearson correlation; lower absolute correlation indicates stronger protection.Source: 2) Fairness · p. 7, 3) Privacy · p. 7

4) Communication overhead:

  • Per-round upload is |w| · 4 bytes per client, and total round cost is 2K|w|·4 bytes for upload plus broadcast.Secure-aggregation overhead is separately reported as key-exchange bytes and client-side compute time.Source: 4) Communication overhead · p. 7

E. Hyperparameter Settings

  • The frequency boundary b=5 assigns IMF1–IMF6 to residual spikes and IMF7–IMF12 to diurnal and seasonal trends.The boundary was selected by validation on one client and fixed across both datasets; tested values were 3, 5, 6, and 7.Source: E. Hyperparameter Settings · p. 7

V. RESULTS AND DISCUSSION

The results assess frequency decomposition, federated aggregation, and DLG resistance. SecFD-FL combines competitive forecast accuracy with client fairness and sharply lower reconstruction correlation under secure aggregation.

A. Effectiveness of Frequency Decoupling

  • CEEMDAN inputs improve recurrent and CNN forecasts on both datasets; on Palo Alto, GRU RMSE falls from 1.44 to 0.80 kWh and R2 rises from 0.33 to 0.79.GRU with CEEMDAN input achieves the best overall results, with RMSE 0.80 kWh and R2 0.791 on Palo Alto, and RMSE 3.55 kWh and R2 0.793 on EV Watt.Source: A. Effectiveness of Frequency Decoupling · p. 7, Table II · p. 8

B. Aggregation Result Comparison

  • Heterogeneity-aware FedProx, SCAFFOLD, and FedDyn perform poorly on at least one client, with SCAFFOLD reaching R2 −0.529 and RMSE 9.82 kWh on EV Watt.The paper attributes this pattern to corrections designed for non-IID classification tasks that can overconstrain updates on stochastic charging signals.Source: B. Aggregation Result Comparison · p. 8
  • SecFD-FL reaches 2.28 kWh RMSE and R2 0.927 on EV Watt, and 0.689 kWh RMSE and R2 0.847 on Palo Alto; its Jain’s fairness indices are 0.871 and 0.991, respectively.On EV Watt, FedAvg has slightly lower RMSE (2.08 vs. 2.28 kWh) and higher R2 (0.938 vs. 0.927), but lacks gradient-inversion protection.Source: B. Aggregation Result Comparison · p. 8, Table III · p. 8
  • The two methods differ by less than 10^-3 in every metric on both datasets when comparing FedAvg-FD without privacy and SecFD-FL with secure aggregation.The zero-sum masks cancel during aggregation; loss decreases smoothly and the global average stabilizes after approximately 25 rounds.Source: B. Aggregation Result Comparison · p. 8, Table III · p. 8, Figure 3 · p. 8

C. Defense Performance against DLG

D. Communication and Compute Overhead

SecFD-FL combines CEEMDAN-driven frequency decoupling, stream-specific GRUs, and zero-sum secure aggregation. The reported forecasting and privacy results come with no added per-round communication and low masking compute overhead.

D. Communication and Compute Overhead

E. Per-Station Forecast Quality

VI. CONCLUSION

  • Across user- and station-level real-world datasets, SecFD-FL matched or exceeded six representative FL algorithms and reduced DLG reconstruction correlation by up to 15.6×.Source: §VI · p. 9
  • Future work will address dynamic client populations with asynchronous participation, differential privacy, and graph-aware aggregation for spatially correlated charging networks.Source: §VI · p. 10

Related papers

Loading 2610.12909v1…