Research on robustness testing and enhancement method for deep learning models
Hailong Deng · IET conference proceedings. · 2025
Currently, deep learning techniques are widely used in safety-critical fields such as autonomous driving and medical diagnosis. However, studies have shown that even advanced deep learning models can be overconfident or make erroneous predictions when encountering perturbed data, which may pose significant safety risks. Ensuring robustness in real-world applications is essential for maintaining model reliability. While prior research has focused on testing model safety, this paper introduces a framework for both testing and enhancing deep learning model robustness. Drawing inspiration from active learning, which has shown that models tend to misclassify high-uncertainty samples, we leverage uncertainty estimation to guide the selection of challenging data points for testing and retraining. By targeting these high-risk samples, we can improve model performance and robustness more effectively. Experimental validation in a safety-critical scenario demonstrates that our approach improves model robustness by an average of 14%, indicating that the capability of the model has been significantly enhanced.