Federated Calibrated-Transfer Learning for Distracted Driving Detection

With the frequent occurrence of traffic accidents around the world, distracted driving detection is crucial for the construction of intelligent transport IoT. Traditional centralised deep learning is limited by data privacy and silo issues, and federated learning, as a distributed machine learning paradigm, provides new ideas, but still faces the problems of data heterogeneity due to the differences in driver behaviours, communication constraints, and limited computational resources of in-vehicle edge devices. To this end, we propose lightweight Federated Calibrated-Transfer Learning (FedCT), which optimises the model performance through a dual mechanism: on the one hand, the introduction of category related marginal parameters dynamically adjusts the decision boundary so that the minority class samples get larger classification intervals, and optimises the clarity of the decision boundary of the feature space through a scaling factor to effectively traditional cross-entropy loss on the classification bias of minority classes; on the other hand, the pretrained backbone network freezing strategy is adopted to reduce the size of trainable parameters and enhance the generalisation ability of the global model with the help of migration learning. Finally, through a large number of experiments, FedCT significantly outperforms the existing federated learning algorithms in terms of classification accuracy and communication efficiency, which provides a theoretically innovative and engineering practical solution for the implementation of federated learning in the field of intelligent driving safety.