Towards driver distraction detection: a privacy-preserving federated learning approach

Driver distraction is the chief reason for road accidents, and the auto-detection system for driver distraction could significantly reduce such misfortune. Sufficient valid samples are the prerequisite for designing a such system, and privacy protection is the restrictive condition to utilize these samples. To address this problem, a privacy-preserving federated learning algorithm for detecting driver distractions during driving is proposed. To prevent gradient leakage of the global model during training and leakage of sensitive data, an adaptive gradient clipping (AGC) mechanism is introduced to protect privacy while reducing the communication load. In addition, a differential privacy technique with a Gaussian mechanism is adopted to further protect the privacy of the participants. Then, a restriction term is added to the objective function to ensure the similarity between the local and global models and to improve the model convergence speed. To overcome the problems of diverse backgrounds and small targets for driver distraction detection, a bounding box loss function SIoU and a self-attention mechanism are used in an improved YOLOv5 detector to promote the model performance. Finally, the effectiveness of the algorithm is validated through experiments on real datasets. Compared with the centralized data training model, the improved YOLOv5 detector effectively improves the performance of driver distraction detection with higher accuracy and robustness.