The widespread deployment of smart meters results in vast amounts of electricity consumption data

which forms the foundation for data-driven electricity theft detection methods. However

as neural networks

commonly used detection models

increase in depth and width

the central cloud infrastructure suffers from significant computational burden. To minimize the number of parameters and reduce computational overhead

this paper introduces an edge computing approach for electricity theft detection utilizing a lightweight multi-channel attention fusion network. The proposed method uses the alternating direction method of multipliers (ADMM) to complete data matrices with missing values. Model training is conducted at the computing center

while edge computing devices detect data collected at the edge. The effectiveness of the proposed model is validated through ablation experiments and comparative studies under various missing data conditions. The results consistently demonstrate that the proposed method excels in all evaluation metrics

confirming its stability and effectiveness compared with other methods. Additionally

hardware experiments are conducted to validate the timeliness and effectiveness of the proposed method on edge devices.