FedSphere-ID: Federated Spherical CNN for Secure Underwater Networks

Monitoring underwater drone networks for intrusion detection presents significant challenges due to dynamic environmental conditions, limited bandwidth, and stringent data privacy requirements. Traditional centralized machine learning approaches, while effective in controlled environments, fall short in addressing critical aspects such as privacy preservation, scalability, and data heterogeneity across distributed nodes. Moreover, existing federated learning (FL) solutions typically rely on uniform model structures and simple aggregation mechanisms, limiting their performance under non-independent and identically distributed (non-IID) data settings. To overcome these limitations, this paper proposes a novel framework Fed Lb SCNN -AOS, which integrates Spherical Convolutional Neural Networks (SCNN) within a federated learning setup using Model-Heterogeneous Aggregation Training (MHAT). In this architecture, each underwater drone trains a personalized SCNN model on its local data, while a central server aggregates soft predictions using an auxiliary SCNN model. The SCNN hyperparameters are further optimized using the Atomic Orbital Search (AOS) algorithm, inspired by quantum electron dynamics, to enhance convergence and detection performance. Experimental evaluations conducted using CIC IDS 2017 and SOLIDS Lab Underwater datasets demonstrate that the proposed framework achieves high intrusion detection accuracy (98.26%), precision (97.90%), recall (97.35%), and F1-score (97.62%), while maintaining low execution time (180 seconds) and memory usage (1.69 GB). These results confirm the framework’s capability to operate efficiently in decentralized, resource-constrained, and privacy-sensitive underwater network environments.