Lightweight TinyML-Enhanced Task Offloading in VANETs for Next-Generation Intelligent Transportation Systems

Abstract

Vehicular Ad Hoc Networks (VANETs) face resource constraints, high node mobility, and stringent latency requirements, especially in safety-critical applications such as collision avoidance, path planning, and emergency braking. Task offloading to nearby vehicles or Roadside Units (RSUs) mitigates local computational limits, but dynamic conditions, unreliable nodes, and rapid topology changes complicate dependable node selection. This paper proposes a Tiny Machine Learning (TinyML)-enhanced, credibility-based task offloading framework for real-time decision-making in vehicular networks. RSUs evaluate vehicle reliability through a three-component Credibility Assessment Module: a Task Assignment Component that distributes lightweight test tasks and filters unreliable nodes via TinyML inference; a Verification Component that applies TinyML anomaly detection to validate execution with minimal cloud dependence; and a Scoring Component that updates credibility using an exponentially weighted delay-penalty model. A hybrid approach then integrates Random Forest (RF) for credibility-aware node selection with TinyLSTM for link stability prediction, while vehicles pre-screen candidates locally by mobility and resource availability before querying RSUs, reducing signalling overhead. Simulations in OMNeT++, Veins, and SUMO using a Luxembourg City mobility trace show a 24% higher task success rate, 9% lower completion time, 21% lower packet drop ratio, and 65% fewer disconnection-induced failures over heuristic, MLP, and deep-RL baselines, at only 9–18% RSU-side CPU overhead. The RF (380 KB) and TinyLSTM (48 KB post-quantization) models fit automotive-grade microcontroller budgets, confirming practical deployability for next-generation Intelligent Transportation Systems (ITS).