TAS-VANET: a trust-aware clustering and misbehavior management architecture for vehicular networks with whale-optimized deep detection

Reliable operation of vehicular ad hoc networks (VANETs) requires more than detecting individual falsified messages: misbehaving vehicles must be identified over time, excluded from privileged roles such as cluster head, and prevented from degrading routing performance. This paper presents TAS-VANET, a trust-aware clustering and misbehavior management architecture that couples (i) a weighted clustering scheme (WCS) for stable cluster-head election, (ii) a per-vehicle trust model updated by detector verdicts, and (iii) a compact supervised sparse autoencoder (SAE) with a skip connection, whose hyperparameters are tuned by the Whale Optimization Algorithm (WOA). The detection component is validated on a public consolidation of the VeReMi Extension benchmark (88,029 records, 17 attack variants) under sender-disjoint 10-fold cross-validation, reaching a mean F1 of 0.8831 (held-out test F1 \(=\) 0.9036); WOA tuning contributes \(+9.34\) F1 points over the default-configured SAE (one-sided Wilcoxon \(p=9.77\times {10}^{-4}\) , paired across identical folds). The integrated system is evaluated in OMNeT + + /SUMO against two trust-based clustering protocols (TACR and QMM-VANET): at 195 vehicles TAS-VANET attains a packet delivery ratio of 89.0%, exceeding TACR by 2.0 percentage points and QMM-VANET by 14.0 percentage points (18.7% relative), with the margin over TACR widening from 2.0 to 3.5 percentage points at 90 km/h; deploying the tuned rather than the default configuration lowers end-to-end delay by 14.3% at the 10% malicious-vehicle ratio. We additionally report two transparent negative findings: strong tabular ensembles remain competitive per-message detectors on this benchmark, and constant-position-offset attacks are essentially undetectable from single-message kinematics alone (AUC \(\approx \) 0.529), which we frame as an open research problem. All reported values trace to released simulation and training artifacts.