Mitigating adversarial attacks and building robust deep learning models for assessing risks in tunnel construction

Deep learning (DL) based models have gained significant attention in the risk assessment of tunnel constructions due to their demonstrated accuracy and effectiveness. Deploying these models in real-world projects raises critical cybersecurity concerns, particularly regarding their susceptibility to adversarial attacks. Thus, this research addresses the following question: Are existing deep learning-based risk assessment models susceptible to attacks, and how can the robustness of these models be improved? To effectively respond to this question, we propose a novel integrated knowledge and data-driven approach to enhance the adversarial robustness of DL models in the risk assessment of tunnel construction. The approach includes: (1) a deep neural network (DNN) based model for risk assessment; (2) a mechanistic model that leverages physical knowledge to generate pseudo-labels; (3) a Knowledge-Enhanced Adversarial Attack (KEAA) algorithm to create adversarial samples; and (4) a hybrid dataset and optimized loss function for updating the DNN model to improve its robustness. The San-yang Road subway tunnel project in Wuhan, China, was used to validate the proposed approach. The results show the effective identification of vulnerabilities in the DNN model and provide a practical solution for enhancing its robustness, thereby improving the defense capabilities of the data-driven approach.