Enhancing IoT Security with Asynchronous Federated Learning for Seismic Inversion

Seismic data/images are critical to understand the structure of subsurfaces. However, for accurate structural analysis, seismic images need to be converted into velocity images that can recognize the depth and thickness of subsurface layers, and this conversion is usually achieved by using seismic full-waveform inversion (FWI). Various Deep neural networks (DNNs) have recently been proposed to replace FWI, and well-trained DNNs typically have lower computational costs but generate similar velocity images as compared to FWI. Yet, training a DNN model is nontrivial, which requires transmitting seismic data that are sensed by seismic receivers in the field to a centralized data center, leading to data privacy and security issues. Seismic field tests are normally conducted in rural areas, where data centers and Internet infrastructures are not available. Hence, it is impossible to train a DNN and achieve seismic inversion in realtime. In this paper, we propose the Asynchronous Federated Learning for Seismic Inversion (AsyncFedInv) framework, which applies multiple lIT devices in terms of edge computing boards to collaboratively train a compact UNet model in realtime based on a novel asynchronous federated learning, where 1) a staleness function is applied to mitigate model staleness, and 2) clients that generate similar local models would suspend its training, thus reducing the communication costs and energy consumption. Simulation results demonstrate that AsyncFedInv achieves a similar convergence rate but lower training loss and better testing performance as compared to a baseline algorithm FedAvg.