Distributed Access Control Method for Trusted Interaction of Robotic Agents in a Decentralized Cyber-Physical Environment

Abstract

The article presents an innovative method for distributed access control of robotic agents in a decentralized cyber-physical system (CPS), which combines an advanced architecture of graph attention neural networks (CAT-GNN) with blockchain technologies. The proposed approach aims to enhance the security, reliability, and fault tolerance of interactions between agents through dynamic behavioral anomaly analysis using CAT-GNN, capable of detecting complex spatio-temporal dependencies in agent behavior. The calculated anomaly score is used for adaptive adjustment of the trust level in agents, directly influencing access decisions to critical resources within the distributed system. Simulation experiments have demonstrated that the CAT-GNN detector outperforms the baseline STAD-GNN model in key metrics such as Accuracy, Fl-score, and ROC-AUC, showing high stability and precision in detecting malicious behavior while varying the number of agents and the proportion of malicious participants. The introduction of a dynamic trust mechanism significantly increased the proportion of successfully completed tasks from 63 to 82 %, while simultaneously reducing errors from over 18 to 8 %. The method relies on the integration of machine learning and distributed ledger protocols, ensuring transparency, immutability, and flexibility in access management. This comprehensive mechanism effectively counters internal and external threats, meeting modern security requirements of industrial and IoT systems. The proposed method is capable of effective scalability and adaptation to changing conditions of cyber-physical systems, confirming its high practical value and promising potential for broad application in critical infrastructures, industry, and transportation networks.