Communication‐Security Co‐Design for Federated Learning in Grant‐Free NOMA IoT Networks

ABSTRACT The rapid growth of 6G Internet of Things (IoT) networks demands scalable and secure learning systems that can support massive device connectivity with minimal coordination overhead. Federated learning (FL) over grant‐free non‐orthogonal multiple access (GF‐NOMA) offers a promising approach by enabling distributed model training with asynchronous uplink access and low signalling cost. However, this setup introduces coupled vulnerabilities: The uncoordinated nature of GF‐NOMA leads to random collisions and residual interference, while the decentralised nature of FL exposes the system to poisoning, Sybil and jamming attacks. These cross‐layer threats jointly degrade model convergence and communication reliability. To address this, we propose Security‐Aware Proximal Policy Optimisation (SA‐PPO), a reinforcement learning framework that co‐designs communication security for FL over GF‐NOMA. SA‐PPO jointly embeds physical‐layer features (e.g., SINR and interference) and learning‐layer signals (e.g., anomaly scores and trust values) into its state, action and reward spaces. This enables the base station to optimise admission control, resource allocation and trust‐weighted aggregation in a unified loop. Unlike prior methods that treat communication and security independently, SA‐PPO learns coordinated strategies that attenuate adversarial impact while preserving update diversity. Simulation results show that SA‐PPO achieves over 90% anomaly detection accuracy, sustains secure participation above 80% and reduces collision‐induced decoding errors by 25% under scenarios with up to 40% compromised devices, while incurring only modest increases in energy and latency. These results demonstrate SA‐PPO's effectiveness for secure, scalable and resilient edge intelligence in future 6G IoT environments.