Federated Reinforcement Learning for Privacy-Preserving Autonomous Decision-Making in Intelligent Edge Networks: A Comprehensive Review

Abstract

ABSTRACT The rapid proliferation of intelligent edge devices, autonomous vehicles, industrial robotics, and Internet of Things (IoT) frameworks has transformed modern computing, necessitating decentralized, real-time decision-making capabilities at the network edge. Traditional centralized Reinforcement Learning (RL) architectures require continuous aggregation of high dimensional state-action data into central servers, exposing edge systems to severe privacy vulnerabilities, bandwidth saturation, and unacceptable latency. Federated Reinforcement Learning (FRL) has emerged as a transformative paradigm that bridges collaborative policy optimization and privacy-preserving distributed computing. By enabling heterogeneous edge nodes to collaboratively train deep policy networks without transmitting sensitive local trajectories or state representations, FRL addresses critical privacy concerns while enhancing decision-making autonomy. This review paper provides a rigorous, in-depth evaluation of state-of-the-art FRL architectures, privacy-preservation mechanisms (including Differential Privacy, Secure Multi-Party Computation, and Homomorphic Encryption), policy aggregation algorithms, and trade-offs between privacy guarantees, convergence efficiency, and communication overhead. We examine application domains across Connected Autonomous Vehicles (CAVs), smart microgrids, edge caching, and robotic swarms. Furthermore, we identify fundamental research gaps regarding dynamic network topology, non-IID environment dynamics, and vulnerability to adversarial manipulation, offering a structured roadmap for future research directions.

KEYWORDS: Federated Reinforcement Learning, Privacy-Preserving Computing, Edge Intelligence, Autonomous Decision-Making, Differential Privacy, Multi-Agent Systems, Communication Efficiency.