TrustGuard-VehNet: An ensemble deep learning framework for trust modelling and malicious node detection in vehicular networks

Abstract

Vehicular ad hoc networks (VANETs) require reliable trust assessment and attack detection to protect safety-critical communication under rapid topology changes. This paper presents TrustGuard-VehNet, an ensemble deep-learning framework that combines a one-dimensional convolutional neural network for short-window behavioral features, a bidirectional long short-term memory network for temporal behavior, and graph attention for relation-dependent neighbor evidence. The fused representation supports continuous trust estimation, malicious-behavior classification, threshold-based isolation, and trust-aware next-hop selection. Within the evaluated controlled simulation, TrustGuard-VehNet achieved 97.8% trust accuracy, a 95.6% attack-detection rate, and 92.4% routing stability. Packet delivery ratio increased from 83.4% to 91.7%, while end-to-end delay decreased by 19.2%. The reported 18.9 ms value is simulation-side detection latency rather than execution time measured on an onboard vehicular processor. These results establish proof-of-concept performance for the reported configuration and motivate validation under broader mobility, channel, density, and hardware conditions.