DP-FedIOD: A Differential Privacy and Federated Learning Based Framework for Aerial Insulators Orientation Detection

Abstract

To address the limitations of deep learning models in detecting aerial insulator images from Unmanned Aerial Vehicles, we present a framework dubbed DP-FedIOD, which utilizes differential privacy and federated learning techniques for identifying the aerial insulators. It tackles the issue posed by current algorithms for inspecting insulators, which employ horizontal anchor frames and are incapable of precisely identifying both insulators and their defective parts. In addition, this study addresses the issue of insulator data being safeguarded by laws and policies that impede its wide-scale collection and dissemination among electric power companies enterprises, resulting in insufficient data volume to train deep learning models. In DP-FedIOD, we have improved the YOLOv5 algorithm by refining its head structure and loss function for directional detection of insulators and their defective parts. Additionally, an attention mechanism module has been incorporated into its backbone to enhance feature extraction capabilities. Furthermore, DP-FedIOD collaboratively trains the global model through federated learning. To prevent privacy leakage in the federated learning process, we also incorporate Laplace noise according to the differential privacy mechanism before uploading the weight information. The experimental outcomes demonstrate that the improved YOLOv5 model attains a [email protected] metric of 95.00%, while DP-FedIOD achieves over 75.10% and 77.90% in precision and recall, respectively. These results indicate that DP-FedIOD has significant practical value in constructing an intelligent grid equipment detection system.

Keywords

• Differential privacy
• federated learning
• insulator
• object detection
• orientation identification