ABSTRACT Addressing the security challenges posed by the deep integration characteristics of industrial IoT and the stringent requirements of edge computing architectures for evaluating model computational efficiency and resource utilization, this study designs and optimizes lightweight deep learning security risk assessment models suitable for edge deployment. Through comparative experiments evaluating the performance of different neural network architectures and constructing an industrial IoT security dataset comprising 60 000 samples, six lightweight backbone networks are systematically compared. Findings indicate Efficient Net achieved the highest accuracy at 96.12%, while Squeeze Net featured the smallest parameter count at 1.25 million pieces and minimal memory usage of 267 MB. Conclusions demonstrate that the designed Reset‐SAE model enables efficient risk assessment at the edge, maintaining 85.67% accuracy even in small‐scale data scenarios. This research provides an effective lightweight solution for industrial IoT edge security protection.
Su et al. (2026) studied this question.