Semantic Explainable Federated Graph Transformer with Digital Twin-Assisted Multi-Agent Intelligence for Secure IoT-Enabled Autonomous Drone Networks

Abstract

The advent of the Internet of Things (IoT) connected autonomous drone networks, new applications are emerging in the fields of smart city, environmental monitoring, precision agriculture, and intelligent aerial surveillance, among others. But, the current UAV systems have several limitations in privacy preservation, dynamic network topology, communication efficiency, real-time decision making, and model interpretability. This paper aims to overcome these drawbacks by presenting a novel framework called Semantic Explainable Federated Graph Transformer with Digital Twin-Assisted Multi-Agent Intelligence (SEFGT-DTMAI) for autonomous drone networks in the context of smart cities. To overcome these limitations, this paper introduces a novel framework, namely Semantic Explainable Federated Graph Transformer with Digital Twin-Assisted Multi-Agent Intelligence (SEFGT-DTMAI) for secure and scalable Autonomous Drone Networks in the smart city domain. The suggested framework includes a Semantic Vision Transformer (ViT) for deriving high-level semantic information from the aerial images, a Dynamic Graph Transformed (DGT) for modelling spatial and temporal interactions between the UAVs, and an Adaptive Federated Learning (AFL) approach to facilitate privacy-preserving collaborative model training without sharing raw data. A Digital Twin continuously updates the virtual swarm of UAVs to the physical swarm, enabling predictive analytics, mission optimization, and proactive fault detection. In addition, a Multi-Agent Intelligence module plans missions, performs edge inference, aggregates data in the cloud, and monitors security, and methods based on Explainable Artificial Intelligence (XAI) such as SHAP, attention visualization, and GNNExplainer are used to offer transparent and trustworthy decision support. The proposed framework is validated through extensive experiments using benchmark UAV and IoT datasets, which prove that the proposed approach outperforms the existing approaches with 99.3% detection accuracy, 99.2% federated learning accuracy, 99.5% packet delivery ratio, and 9 ms end-to-end latency, while also reducing the communication overhead and improving the synchronization of Digital Twins. Additional validation of the contribution of each module is provided by ablation studies and statistical analyses. The framework proposed in this paper offers a strong, privacy-preserving and explainable solution for the next generation of intelligent UAV ecosystems in dynamic IoT environments.