Mainstream existing driver identification methods rely on triplet-loss-trained embedding models, whose discriminative performance degrades rapidly as the number of drivers grows and which are prone to overfitting to specific driving session patterns under strict temporal evaluation.