Federated Learning in Underwater Sensor Networks

Abstract

Underwater sensor networks (UWSNs) are essential for environmental monitoring, resource exploitation, disaster control, and maritime safety. Usage of machine learning (ML) algorithms within the networks, however, is not applicable due to high latency, low bandwidth, frequent topological changes, and privacy-stringent aquatic environments: Principles of the Union: Autonomous Acoustic and Hybrid Networks. We propose a distributed learning model that enables the training of models on remote nodes. However, the exchange of model information rather than raw data minimizes the associated overhead of transmission and risk to privacy. In this chapter, we present the first systematic consolidation of federated learning (FL)-integrated UWSNs prototypes designed for secure marine surveillance and other critical utilities. We carefully parallel them with UWSNs’ restrictions and other challenges beyond those determining FL. This process leads us to describe architectural reconfigurations, communication approaches, and workflow redesigns that make FL realistic on acoustic and hybrid principles alike. We shall also provide real-world application mini-cases in environmental monitoring, aquaculture, maritime defense, disaster response, and oceanographic research. We then identify key bottlenecks, such as communication overhead, energy constraints, non-IID training data, and adversarial players, and suggest corresponding solutions, including model compression, adaptive scheduling learning, and secure aggregation, along with lightweight privacy solutions. Lastly, we shall outline the future strategy linking FL with blockchain-driven orchestration, edge processing, digital twins, and standard benchmarks to expedite the experimental evaluation. Ultimately, FL is positioned as the foundational blueprint for safe, robust, and adjustable underwater sensing in marine intelligence systems.