PrivEdge: a hybrid split–federated learning framework for real-time electricity theft detection on edge nodes

TL;DR

PrivEdge is a hybrid machine learning framework that detects electricity theft in real-time on edge devices while preserving privacy and reducing communication costs.

Contribution

PrivEdge introduces a novel hybrid Split–Federated Learning framework for privacy-preserving, real-time electricity theft detection on edge nodes.

Findings

PrivEdge achieves better detection accuracy and F1-score than centralized and standalone Split or Federated Learning baselines, especially under non-IID data conditions.

The framework demonstrates low inference time and consistent resource consumption on Raspberry Pi 4 hardware.

Privacy-preserving techniques like secure aggregation and homomorphic encryption are effectively integrated without compromising performance.

Abstract

Electricity theft is one of the primary contributors of non-technical losses in contemporary power grids, and traditional centralized methods of detection are limited in scale, feature a large communication cost, and create privacy issues. The presented paper introduces PrivEdge, a deployment-friendly hybrid Split–Federated Learning (SL–FL) system to detect real-time electricity theft on resource-constrained edge devices. PrivatEdge uses a Raspberry Pi 4-based smart meter gateway to do localized preprocessing with the Raspberry Pi 4 smart meter gateway and run a lightweight LSTM-based FrontNet; server-side functionality does more in-depth model inference, collaborative coordination, ensemble stacking, and score-level fusion. Split Learning allows conveying small intermediate activations as opposed to raw consumption data, which significantly lowers communication costs and minimizes privacy loss. Federated Learning supports distributed learning between highly non-IID clients who are geographically well-spread. Privacy maintenance is realized by secure aggregation and Laplace differential privacy, where ε = 3 is used as a uniform operation compromise due to practical consideration. As a high-security deployment mode, homomorphic encryption is supported. Extensive experiments on the SGCC smart meter data with IID and non-IID conditions reveal that PrivEdge would perform better in terms of detection accuracy and F1-score than both centralized and FL-only or SL-only baseline frameworks, especially in non-IID conditions. The software-level assessment using Raspberry Pi 4 hardware establishes a low inference time, consistent resource consumption, and endurance at that rate using sustained load. Ablation experiments also confirm the importance of localized preprocessing, time expression, ensemble-based aggregation of data, and their privacy-conscious learning. In general, PrivEdge helps in closing the gap between hybrid concepts of SL–FL learning and the practical needs of deployment in privacy-aware electricity theft detection at the network edge.

Figures