Abstract:“Safety is our friend, accidents our enemy” is the propaganda social information on traffic safety in several countries. This proves that safety is a top priority when using vehicles when participating in traffic. There have been several studies on driver behavior/state detection using machine learning (ML). In addition, in ML generally, the data must be provided to the server for training in a centralized setting. Sharing data will therefore expose data information to external interference and corruption, bringing many risks. Federated learning (FL) recently emerged as new research to solve the tricky problems above. Our research is one of the few first studies that applies longitudinal physiological sensor measurements in a naturalistic driving environment using the FL framework to assess driver behavior/state. The novelty of this study includes the following. First, we use techniques such as synthetic minority oversampling technique (SMOTE) to enrich and balance the Harmony dataset. The purpose is to increase the accuracy and improve the reliability of the assessment. Second, we incorporate essential deep learning (DL) models into the FL framework, including multilayer perceptron (MLP), long short-term memory (LSTM), 1-D convolutional neural network (1-D CNN), CNN-LSTM, and convolutional LSTM (ConvLSTM) models. By evaluating these models, we show that in the FL environment, the use of compact, lightweight models such as MLP provides higher efficiency than other large, complex models and is suitable for consistent research. Finally, our framework has higher results than other basic models such as support vector machines and decision trees while remaining highly competitive to other DL models with respect to network size, execution time, and stability. Primarily, our idea is the trend of the future. It can be practically applied in transportation, taxis, and self-driving car companies; each car is considered a client, and companies can manage their drivers instead of using a third-party service, which will cause data loss and privacy.

FedDAF: Federated Deep Attention Fusion for Dangerous Driving Behavior Detection

SLAFusion: Attention fusion based on SAX and LSTM for dangerous driving behavior detection.

Symbolic Aggregate Approximation Based Data Fusion Model for Dangerous Driving Behavior Detection

Video-Based Abnormal Driving Behavior Detection Via Deep Learning Fusions

FDAN: Fuzzy deep attention networks for driver behavior recognition

RFDANet: an FMCW and TOF radar fusion approach for driver activity recognition using multi-level attention based CNN and LSTM network

Safety Is Our Friend: A Federated Learning Framework Toward Driver’s State and Behavior Detection

Temporal Information Fusion Network for Driving Behavior Prediction.

A Multimodal Data-Driven Approach for Driving Risk Assessment

MM-LMF: A Low-Rank Multimodal Fusion Dangerous Driving Behavior Recognition Method Based on FMCW Signals

mm-DSF: A Method for Identifying Dangerous Driving Behaviors Based on the Lateral Fusion of Micro-Doppler Features Combined

Deep Attentional Factorization Machines Learning Approach for Driving Safety Risk Prediction

Expressway Traffic Incident Detection Using a Deep Learning Approach Based on Spatiotemporal Features with Multilevel Fusion

STDA: Spatio-Temporal Dual-Encoder Network Incorporating Driver Attention to Predict Driver Behaviors Under Safety-Critical Scenarios

FedBiKD: Federated Bidirectional Knowledge Distillation for Distracted Driving Detection

Peer-to-Peer Federated Continual Learning for Naturalistic Driving Action Recognition

Federated Transfer-Ordered-Personalized Learning for Driver Monitoring Application

DADA: Driver Attention Prediction in Driving Accident Scenarios

Deep Federated Learning for Autonomous Driving

AFM3D: An Asynchronous Federated Meta-Learning Framework for Driver Distraction Detection