Robust federated learning for UAV object detection: a joint self-distillation and drift compensation approach

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned aerial vehicles (UAVs) are crucial for disaster response and environmental monitoring.
• Real-time object detection in UAV swarm networks faces challenges due to non-independent and identically distributed (non-IID) data.
• Model convergence and adaptability are hindered by data heterogeneity in UAV networks.

Purpose of the Study:

• To develop a robust federated UAV object detection framework for non-IID data.
• To enhance client adaptability, detection performance, and convergence speed.
• To address model drift and improve synchronization between local and global model updates.

Main Methods:

• Implemented a federated UAV object detection framework.
• Incorporated a self-distillation mechanism using historical local model states for guided training.
• Introduced a drift compensation mechanism to synchronize local and global model updates.
• Conducted experiments on the VisDrone2019-DET dataset.

Main Results:

• The proposed framework accelerated convergence speed by approximately 2.2 times.
• Achieved an approximate 3% enhancement in detection performance.
• Demonstrated improved adaptability and robustness under non-IID data conditions.
• Outperformed baseline models in experimental comparisons.

Conclusions:

• The developed framework offers an efficient and robust solution for UAV-based object detection in non-IID scenarios.
• Self-distillation and drift compensation effectively improve federated learning performance for UAV swarms.
• The approach balances model specialization and adaptability, crucial for dynamic environments.