Neural Network Repair With Shapley‐Guided Search

ABSTRACT The deployment of deep neural networks (DNNs) in safety‐critical domains is critically hampered by their vulnerability to defects, which can arise from malicious attacks or low‐quality data. Therefore, precisely locating the network components responsible for these defects, and subsequently repairing them without compromising overall model performance, presents a significant challenge. To address this, this paper introduces NSRepair, a framework that combines interpretable fault localisation with multi‐objective optimisation. Specifically, to accurately attribute blame for a defect, we employ Shapley values to quantify the contribution of each neuron. To systematically manage the trade‐off between defect correction and performance preservation, we formulate the repair task as a multi‐objective optimisation problem. We conducted extensive experiments across four distinct repair tasks, validating NSRepair on diverse model architectures against seven specialised state‐of‐the‐art methods. The results demonstrate that our unified framework effectively repairs a wide range of defects, demonstrating its potential as a versatile and practical solution for improving DNN dependability. Our code is publicly available at https://doi.org/10.5281/zenodo.17494304 .