Pareto optimized neural architecture search framework for edge based driver distraction detection

Abstract The growing adoption of Artificial Intelligence (AI) in real-world applications highlights the need for Deep Learning (DL) models that are both efficient and accurate. However, the high computational requirements of DL often hinder deployment in resource-limited environments. In this paper, we propose an efficient and resource-conscious extension of Differentiable Architecture Search (DARTS) to design lightweight neural networks for driver distraction detection. The proposed method introduces a multi-objective optimization mechanism grounded in Pareto efficiency. It explicitly balances accuracy, latency, and model size to enable deployment-aware architecture search. Experiments conducted on the State Farm Dataset (SFD) and the American University in Cairo Dataset (AUCD2) demonstrate up to a 7 $$\times $$ reduction in parameters compared to state-of-the-art NAS methods while preserving high accuracy ( $$98.17\%$$ and $$95.80\%$$ , respectively, with only $$-1.7\%$$ and $$-0.98\%$$ accuracy trade-offs). The optimized models, with sizes as small as 0.25 Megabytes (MB) and 0.36 MB, achieve real-time inference latencies of 3–4 milliseconds on an Nvidia Jetson Xavier NX, significantly improving search efficiency and edge deployability.