Maritime Decentralized Graph Federated Learning Algorithm with Error Compensation

Based on the characteristics of Graph Neural Network (GNN) processing graph data, through the calculation and prediction of the relationship between the change of marine environmental values over time and the activity trajectories of marine organisms, the dynamic activities of marine organisms can be better grasped, which is helpful for the protection of marine biodiversity. However, due to environmental, geographic location, privacy and other factors, the development of maritime communication technology is obviously lagging behind that of land. Federated learning (FL) has been shown to have an outstanding effect on the protection of data privacy,but the exploration of applying federated learning in resource limited marine communication environments is still far from sufficient. And it also pushes forward the development of graph neural networks due to the existence of some kind of connection between clients in federated learning. Therefore, we propose a DeFedGNN algorithm that achieves a trade-off between computation and resource consumption while ensuring privacy and model accuracy.