Building Trust for Underwater Wireless Sensor Networks via Riemannian Variational Autoencoders

With sensitive applications in marine exploration, disaster warning, and oil monitoring, ensuring the data integrity of Underwater Wireless Sensor Networks (UWSNs) has become an increasingly critical challenge. Existing trust models primarily rely on euclidean distance to evaluate the trustworthiness of nodes. However, the lack of a probabilistic framework for representing trust evidence undermines the accuracy and reliability of these models. In this paper, we propose a novel trust model based on a Variational Autoencoder (VAE) that maps trust evidence into a latent space to model the normal data distribution, thus enabling the identification of malicious nodes. Recognizing the limitations of euclidean latent spaces in capturing the non-linear characteristics of underwater acoustic signals, we extend this framework by introducing a Riemannian manifold as the latent space. The proposed method enhances scalability by reconstructing implicit features, thereby offering the potential to uncover previously unrecognized threats and improving the accuracy of malicious node identification. Simulation results demonstrate that the proposed trust model achieves a detection accuracy of 94%. Compared to existing trust models, this approach exhibits superior verification accuracy and robustness, making it a promising solution for ensuring the reliability of UWSNs in underwater environments characterized by high noise levels and dynamic topology changes.