Attention in Motion (AIMformer): Transformer-Based Misbehavior Detection for Securing Vehicle Platoons

AIMformer is a Transformer-based real-time misbehavior detection framework for vehicle platoons, targeting attacks in which authenticated vehicles inject forged kinematic data. The framework uses multi-head self-attention to capture intra-vehicle temporal dynamics, a spatiotemporal variant to model inter-vehicle spatial correlations, and a PFBCE loss function that penalizes false positives to suit safety-critical systems. Across 4 platoon controllers, multiple attack vectors and diverse mobility scenarios, performance is no lower than 0.93 and outperforms SOTA baselines; deployments based on TFLite, ONNX and TensorRT achieve sub-millisecond inference latency, suitable for in-vehicle and roadside edge platforms.