SMUSAC: Lightweight federated learning framework for SUNETs with tolerance of data loss and node compromise

Abstract

Satellite-Assisted Unmanned-System Networks (SUNETs) are emerging network applications that leverage satellites to support ubiquitous data-driven services, such as autonomous underwater vehicles and unmanned aircraft systems. In these applications, transmitting data over external networks poses a risk of privacy leakage. Usually, federated learning is used to prevent the direct leakage of raw data; however, its effectiveness and robustness in SUNETs are constrained due to two key challenges arising from limited bandwidth and unmanned nodes: (a) data loss, some nodes may fail to transmit data back to the server in time; (b) node compromise, unmanned nodes might be controlled by adversaries, even uploading malicious data to the server. To address these challenges, we propose a lightweight federated learning framework, called SMUSAC, which includes three stages: Sparsifying Model, Uploading Signs, and Aggregating with Compensation. Specifically, we design a sign-based updating mechanism for sparsified models, rather than transmitting model parameters or gradients over the communication link. It improves SMUSAC’s tolerance to data loss and node compromise by relying solely on the sign of updates rather than their specific values, while also reducing bandwidth demands. Additionally, an error-compensation mechanism is employed to mitigate the accuracy loss caused by sparsification. We theoretically analyze the convergence of SMUSAC under a non-convex cost function. Simulation results show that SMUSAC exhibits significant resilience under adverse conditions, maintaining stable performance even with 40% of nodes compromised, and outperforms seven baselines across multiple evaluation metrics.