Enhancing adversarial robustness in driver drowsiness detection using dual-teacher guided adversarial distillation (DTGAD)

Drowsy driving, a major road safety issue requires reliable detection. Deep learning-based driver monitoring systems leveraging computer vision have shown considerable potential for automated drowsiness detection using visual and behavioral cues. However, their reliability under adversarial conditions such as evasion attacks remains a critical concern, where an incorrect drowsiness prediction can have serious consequences. Existing adversarially robust learning and knowledge distillation approaches improve robustness, but effectively transferring accuracy as well as adversarial robustness to a compact student model remains challenging. This research work proposes a novel Dual-Teacher Guided Adversarial Distillation (DTGAD), a dual teacher framework with an adaptive weight mixing strategy to enhance adversarial robustness by combining complementary knowledge of a clean teacher and a robust teacher. The contribution of each teacher is dynamically adjusted according to its prediction confidence for individual samples. Proposed DTGAD framework is evaluated on CIFAR-10 benchmark and image-based drowsiness datasets,including Driver Drowsiness Detection (DDD) extracted from Real-Life Drowsiness Dataset (RLDD), Closed Eyes in Wild (CEW) dataset and Eye-Blink dataset. Performance of proposed DTGAD is compared against existing adversarial distillation defenses, demonstrating its usefulness for reliable drowsiness detection under adversarial conditions.