METHOD OF DYNAMIC TRUST ASSESSMENT IN DISTRIBUTED SYSTEMS

Abstract

Context. Modern distributed computing and cyber-physical systems, such as UAV swarms and IoT networks, are shifting from centralized control to decentralized real-time collective decision-making. These systems often operate under conditions of limited bandwidth, dynamic topology, and the presence of unreliable or malicious (“Byzantine”) nodes.Objective. The goal of this work is to increase the security and efficiency of distributed systems by developing an adaptive mechanism for trust evaluation that can operate in uncertain environments.Method. The study proposes a Method of Dynamic Trust Assessment that utilizes a Recurrent Neural Network (LSTM) to process time-series data of node behavior. Unlike static models, this approach implements an online learning mechanism to continuously predict adaptive trust thresholds based on the consistency of node actions with expected trajectories and groupconsensus. The method evaluates multi-factor features, including the coherence of observations and the regularity of message exchange, to calculate a dynamic reputation score for each participant. Furthermore, the system employs an adaptive validation procedure for collective decisions, automatically adjusting the required trust level in response to detected anomalies. This allows the distributed system to transition seamlessly between high-efficiency modes and high-security modes depending on the current threat landscape. By integrating these components, the method effectively neutralizes the influence of “Byzantine” nodes without disrupting the overall system coordination.Results. Experiments conducted in a ROS 2 and Gazebo simulation environment confirmed the effectiveness of the proposedmethod. The neural network-based approach improved the accuracy of threat detection at individual nodes (precision increased by 4.3% and recall by 7.8%) compared to the baseline Trust-WCA method. Furthermore, the overall effectiveness of the system in executing collective tasks – measured by the ratio of successful target interactions versus false targets – improved by 8%, with a corresponding increase in the F1-Score by 5.9%.Conclusions. The proposed adaptive method ensures increased stability of distributed systems against cyber attacks and “Byzantine” faults without significant computational overhead, making it suitable for real-time edge computing applications