OES-Fed: a federated learning framework in vehicular network based on noise data filtering.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Federated Learning

Background:

• The Internet of Vehicles (IoV) faces challenges with insufficient data sharing and noisy sensor data impacting machine learning (ML) model accuracy.
• Vehicle data reluctance and in-vehicle camera issues like shaking and obscuration hinder effective ML model development in IoV.

Purpose of the Study:

• To propose a novel federated learning framework, OES-Fed, to address data scarcity and noise in the IoV.
• To enhance the accuracy and efficiency of ML models within the IoV environment.

Main Methods:

• Developed the Outlier Detection and Exponential Smoothing federated learning (OES-Fed) framework.
• Implemented noise data filtering using a combination of data outlier detection, K-means clustering, Kalman filtering, and exponential smoothing algorithms.
• Evaluated the framework from both current and historical data perspectives for local ML models.

Main Results:

• The OES-Fed framework demonstrated superior performance across three datasets.
• Achieved higher accuracy, reduced loss, and improved Area Under the Curve (AUC) compared to existing methods.
• Successfully filtered noise data, enhancing the reliability of ML models in IoV.

Conclusions:

• The OES-Fed framework effectively filters noise data in the IoV, overcoming limitations of data sharing and sensor quality.
• OES-Fed provides a valuable reference for implementing federated learning in IoV applications.
• The proposed method significantly improves ML model accuracy and communication efficiency in intelligent vehicles.