Privacy Preserving Federated Reinforcement Imitation Learning Framework With Robust Aggregation for Cloud‐Based Domestic Robots

摘要

ABSTRACT The evolution of cloud‐based web services for the control and tracking of robotic devices has significantly transformed the field of robotics. Domestic robotic devices performing household activities are becoming increasingly popular, as they collect large volumes of data, transmit it to the cloud, and leverage web services for collaborative learning. These devices interconnect and learn from their peers over the cloud. However, this distributed and interconnected learning environment introduces a serious vulnerability to model poisoning attacks, where malicious participants can deliberately corrupt the learning process. These attacks are a direct result of Byzantine behavior, where certain participants act arbitrarily or adversarially, undermining the integrity of the global model. These attacks pose a critical threat to the reliability, safety, and privacy of robotic systems operating in real‐world environments. To accelerate learning, peers connected through cloud‐based services contribute data and updates, but this collaboration inevitably leads to the exposure of sensitive information, further escalating privacy concerns. To tackle these pressing issues, we propose a novel framework called Federated Reinforcement Imitation Learning (FRIL). The framework involves the design of the FRIL architecture, an in‐depth analysis of threats in a distributed setting, and the development of a robust algorithm specifically designed to defend against model poisoning attacks. Experimental results demonstrate a high learning accuracy of 88 percent using the Edge IIoT dataset. The collaborative, decentralized, and privacy‐preserving nature of the proposed framework, combined with imitation learning, makes it highly resilient against adversarial interference, ensuring the stability and integrity of the Federated Learning process in domestic robotic environments. This work directly targets the growing threat of model poisoning attacks and provides a concrete solution to secure collaborative learning in intelligent robotic systems.