Distracted　driving　is　a　significant　cause　of　traffic　accidents　and　has　becoming　an　increasingly　severe　road　safety　concern,　in　particular　with　driver　distractions　such　as　using　phones,　talking,　and　makeup　applications.　Therefore,　real-time　detection　of　driver　behavior　and　reactions　during　driving　is　crucial　for　improving　traffic　safety.　To　resolve　the　problems　of　large　parameter　counts　and　inaccurate　classification　in　existing　distracted　driving　detection　models,　this　paper　presents　a　lightweight　model,　named　RepViT-MCFM,　which　features　low　latency　and　robustly　detects　driver　distraction.　Specifically,　we　designed　a　multi-scale　cross-fusion　denoising　module　(MCFM)　based　on　Haar　wavelet　transform,　utilized　the　RepViT　module　improved　from　MobileNetV3　as　the　backbone　feature　extractor,　and　designed　the　RepViT-MCFM　model　with　a　low-latency　simple　classifier.　We　have　conducted　quantitative　and　qualitative　evaluations　on　the　State-Farm　dataset,　extensive　experiments　demonstrate　that　the　designed　RepViT　model,　compared　to　existing　CNN-based　detection　models,　maintains　high　accuracy　while　achieving　lower　latency　and　fewer　parameters.　©　2024　IEEE.