A Collision-Aware Optimization Framework for Wireless Federated Learning over Grant-Free NOMA

This paper proposes a novel Collision-Aware Successive Convex Optimization (CA-SCO) framework for wireless federated learning (WFL) over grant-free non-orthogonal multiple access (GF-NOMA). Existing WFL systems approaches predominantly adopt grant-based NOMA or orthogonal multiple access schemes, which inherently incur scheduling overhead and limit scalability. GF-NOMA enables scalable and asynchronous transmissions without scheduling overhead; it introduces challenges such as stochastic collisions, residual successive interference cancellation (SIC) errors, and retransmission, which impair model convergence and energy efficiency. To address these limitations, we present a unified optimization framework that jointly models probabilistic collision dynamics, residual SIC interference, and retransmission-induced delays within GF-NOMA-enabled FL. A joint latency–energy minimization problem is formulated under signal-to-interference-plus-noise ratio (SINR) and reliability constraints, taking into account practical device limitations and imperfect channel state information (CSI). To solve the resulting non-convex problem, we develop a provably convergent successive convex approximation (SCA) algorithm that incorporates surrogate modeling, trust-region adaptation, and SIC-aware decoding order updates. Simulation results involving up to 200 edge devices show that CA-SCO reduces latency by up to 35%, improves energy efficiency, and accelerates model convergence compared to state-of-the-art baselines. These findings establish CA-SCO as an effective and scalable solution for ultra-reliable, low-latency edge communications in beyond-5G edge networks.