AI-DRIVEN FEDERATED MULTI-AGENT ACTOR–CRITIC LEARNING FORSECUREANDENERGY-EFFICIENT RESOURCE OPTIMIZATION IN 6GSEMI-GRANT-FREE NOMA-BASED IOT NETWORKS

Sixth-generation (6G) wireless systems are expected to support ultra-massive machinetype communication through dense deployments of Internet of Things (IoT) devices.Semi-Grant-Free Non-Orthogonal Multiple Access (SGF-NOMA) has emerged as apromising access mechanism for enabling simultaneous transmissions of grant-basedand grant-free users while improving spectral efficiency. Dense IoT environmentscreate significant challenges including power-domain collisions, energy inefficiency dueto repeated retransmissions, and susceptibility to malicious interference such asjamming attacks. Conventional centralized optimization techniques introduce highsignaling overhead and raise privacy concerns, limiting their scalability in large-scaleIoT deployments. This work proposes an AI-driven federated hybrid multi-agent actorcritic learning framework for secure and energy-aware resource optimization in 6GSGF-NOMA IoT networks. Each grant-free device operates as an intelligent learningagent that autonomously selects transmission power levels and resource blocks basedon local observations. A federated learning architecture enables decentralized modeltraining while periodically aggregating parameters at an edge server using federatedaveraging. The proposed framework integrates a hybrid exploration–cooperationstrategy and a multi-objective reward function that jointly considers throughput gain,collision penalties, energy consumption cost, and security robustness under jamminginterference. Simulation-based evaluations demonstrate that the proposed frameworksignificantly improves conditional throughput, reduces power collision probability, andenhances energy efficiency compared with centralized deep reinforcement learning,random access, and conventional SGF-NOMA approaches. Results also indicatefaster convergence and improved fairness among IoT devices under ultra-densedeployment scenarios. The proposed solution provides a scalable and privacy-preservinglearning architecture suitable for AI-native 6G wireless systems and large-scale IoTnetworks.