Argus: Federated Non-convex Bilevel Learning over 6G Space-Air-Ground Integrated Network

TL;DR

Argus addresses decentralized, non-convex bilevel learning over time-varying SAGIN by enabling asynchronous, serverless collaboration among heterogeneous agents. It converts the bilevel problem to a single-level form using a lower-level estimation step and a polyhedral outer approximation built from cutting planes, accommodating non-smooth objectives via proximal updates. The authors establish convergence with an iteration complexity of $\mathcal{O}(1/\epsilon)$ and derive corresponding communication and computational costs, supported by rigorous lemmas and proofs. Empirical results across meta-learning, hyperparameter optimization, and continual learning demonstrate that Argus outperforms synchronous baselines and remains robust under stragglers, highlighting its practical impact for scalable 6G SAGIN deployments.

Abstract

The space-air-ground integrated network (SAGIN) has recently emerged as a core element in the 6G networks. However, traditional centralized and synchronous optimization algorithms are unsuitable for SAGIN due to infrastructureless and time-varying environments. This paper aims to develop a novel Asynchronous algorithm a.k.a. Argus for tackling non-convex and non-smooth decentralized federated bilevel learning over SAGIN. The proposed algorithm allows networked agents (e.g. autonomous aerial vehicles) to tackle bilevel learning problems in time-varying networks asynchronously, thereby averting stragglers from impeding the overall training speed. We provide a theoretical analysis of the iteration complexity, communication complexity, and computational complexity of Argus. Its effectiveness is further demonstrated through numerical experiments.

Figures (11)

• Figure 1: The architecture of decentralized federated learning framework in SAGIN.
• Figure 2: Convergence performance of Argus.
• Figure 3: Performance comparison with baseline methods on meta-learning.
• Figure 4: Performance comparison with baseline methods on hyperparameter optimization.
• Figure 5: Performance comparison with baseline methods on continual learning.

Theorems & Definitions (3)

• proof
• proof
• proof