Driver Action Recognition Using Federated Learning

Abstract

Distracted driving action is the main cause of car accidents. In recent years, CNN-based driving action recognition methods have become mainstream. However, these methods train CNN model using a centralized manner, not only need to collect a large amount of data in advance, which may leak user privacy, but also make it difficult for model upgrades. In this paper, we use federated learning for model training, which protects user privacy while achieving online model upgrades. The experiments based on the State Farm dataset show that whether it is under iid or non-iid settings, the accuracy achieved by the model trained using federated learning is competitive with that of the model obtained by centralized training.