FedOTO: Toward Low-Latency Drowsiness Detection Service in IoV via Structured Federated Pruning

Abstract

Drowsy driving continues to be a leading cause of motor vehicle accidents (MVAs), accounting for nearly 30% of all incidents according to recent studies. While numerous drowsiness detection methods have been developed, existing solutions often struggle with practical deployment due to limitations in detection accuracy, processing latency, or computational efficiency. To overcome these challenges, we propose federated only-train-once (FedOTO), a low-latency drowsiness detection framework for the internet of vehicles (IoV). FedOTO integrates edge computing with a novel two-tier structured federated pruning framework to deliver personalized, accurate, and real-time drowsiness detection services. By leveraging individual user data, local models on vehicular edge nodes are dynamically pruned and optimized to capture the individual drowsiness patterns. To reduce the processing latency, we employ zero-invariant groups (ZIGs), which identify computation-redundant parameter groups that can be safely pruned, and a hybrid efficient structured sparse optimizer (HESSO), which automatically trains local models and performs pruning, thereby minimizing local training time, communication delay, and response time. Additionally, our co-adaptive strategy dynamically adjusts both the number of redundant groups and learning rate across iterations, thus significantly accelerating the model convergence. Extensive experiments demonstrate FedOTO’s superior performance, achieving an average accuracy of 0.9997 and an average F1 score of 0.9989. Moreover, compared to conventional federated learning, FedOTO reduces local training time by 30.7%, communication delay by 28.9%, and response time by 34.3% to achieve superior overall performance over state-of-the-art baselines.