Octopus: Knapsack model-driven federated learning client selection in internet of vehicles

Federated learning (FL), as a distributed way for processing real-time vehicle data, is widely used to improve driving experience and enhance service quality in Internet of Vehicles (IoV). However, considering the data and devices heterogeneity of vehicle nodes, randomly selecting vehicles that are involved in model training would suffer from data skewness, high resource consumption, and low convergence speed. To this end, we propose Octopus, which consists of two components: i) an importance sampling-based local loss computation method is designed to request resource information for each client and apply the importance sampling technique to assess each client’s contribution to the global model’s convergence, followed by utilizing a knapsack model that treats the local loss of each client as the item value, while treating the total system training time as the knapsack capacity to accelerate the client convergence; ii) a knapsack model-based federated learning client selection method is designed to select the client with optimal local loss and maximum model uploading speed to participate in training. In each training round, these clients download and update the model within a predefined time, followed by enabling the selected clients to continue uploading the updated model parameters for assisting the server to efficiently complete the model aggregation. Experimental results show that Octopus improved the model accuracy by 2.64% $\sim$32.61% with heterogeneous data, and by 1.97% $\sim$11.74% with device heterogeneity, compared to eight state-of-the-art baselines.