Statistical Federated Learning for Beyond 5G SLA-Constrained RAN Slicing
Hatim Chergui, Luis Blanco, Christos Verikoukis
Abstract
A key enabler for both scalability and sustainability in beyond 5G (B5G) network slicing consists on minimizing the exchange of raw monitoring data across different domains. This is achieved by bringing the analysis functions closer to the data collection points. To this end, we introduce in this paper <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">statistical federated learning</i> (SFL) provisioning models that can learn over a live network non independent identically distributed (non-IID) datasets in an offline fashion while respecting slice-level service level agreement (SLA) long-term statistical constraints. Specifically, we consider three resource SLA metrics, namely, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">cumulative distribution function</i> (CDF), <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$Q$ </tex-math></inline-formula> - <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th percentile</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">maximum/minimum bounds</i> . These metrics are dataset-dependent and non-convex non-differentiable and, to sidestep the inaccuracy of settling only for surrogates, we propose a novel formulation that jointly considers the statistical objective and constraints as well as their smooth approximation using the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">proxy-Lagrangian</i> framework, which we solve via a non-zero sum two-player game strategy. Numerical results on various slice-level resources show that SFL enables SLA enforcement while significantly reducing the overhead compared to both state-of-the-art FedAvg and centralized constrained deep learning schemes. Finally, we provide an analysis for the lower bound of the so-called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">reliable convergence probability</i> in the SFL setup.