Hybrid LSTM-GNN-Driven Blockchain Framework for Secure IoT Device Authentication and Predictive Data Intelligence

Abstract

This study proposes a hybrid LSTM-GNN-driven blockchain framework for secure IoT device authentication and predictive data intelligence. The framework integrates a Long Short-Term Memory (LSTM) network for temporal anomaly detection, a Graph Neural Network (GNN) for structural anomaly detection, and a blockchain-based trust-management layer that records device identity and communication activity through smart contracts. As the primary contribution of this work, the proposed framework is evaluated independently on a simulated IoT testbed of 500 heterogeneous devices: the LSTM branch achieves 97.0% temporal anomaly detection accuracy (MAE 0.147, RMSE 0.212), the GNN branch achieves a structural anomaly F1-score of 88.9%, and the blockchain trust-management layer sustains a mean latency of 49.9 ms across all 500 simulated devices. To validate the discriminative strength of the underlying feature-engineering pipeline, five classical machine-learning classifiers K-Nearest Neighbors (KNN), Decision Tree (DT), Gradient Boosting (GB), Extra Trees (ET), and Random Forest (RF) are additionally benchmarked on the NSL-KDD dataset using ExtraTreesClassifier-based feature selection, SMOTE oversampling, and 5-fold cross-validation, with Random Forest attaining the highest binary classification accuracy of 99.92% (F1 = 99.92%) and multi-class accuracy of 99.93%. These NSL-KDD results serve solely as supporting evidence for the engineered feature-preprocessing pipeline and do not, by themselves, validate the proposed hybrid architecture; the framework-level LSTM, GNN, and blockchain metrics remain the central evidence for the paper's claimed contribution. The findings demonstrate that combining deep learning-based behavioral analysis with blockchain-based decentralized trust management provides a scalable and trustworthy foundation for future IoT security systems.