Enhancing robustness of deep learning models against adversarial attacks for construction worker safety

The increasing use of deep neural networks (DNNs) in construction safety systems highlights their potential but also reveals vulnerabilities to adversarial perturbations. Such weaknesses can lead to false detections, increasing the risk of accidents on dynamic construction sites. This paper advances construction safety research by developing a framework to enhance AI robustness through adversarial training (AT) using the TRADES method with $L_{\infty }$ and $L_2$ norms on a ResNet-18 architecture. The approach was evaluated using a combined dataset of publicly available construction images and custom-collected lab data representing unsafe behaviors. Results show the adversarially trained model achieved 92.50 % benign accuracy and 90.36 % robust accuracy under L₂ attacks. To assess model transparency, LIME (Local Interpretable Model-Agnostic Explanations) was used to visualize regions influencing predictions for both benign and adversarial inputs. These improvements support safer, AI-assisted monitoring in real-world settings by enabling more reliable decision-making and reducing the risk of AI system failures.