Filter-based Repair of Semantic Segmentation in Safety-Critical Systems

Deep Neural Networks (DNNs) have become a core component in several safety-critical tasks like Autonomous Driving. To ensure their safe usage in these scenarios, DNN repair approaches have been used to fix erroneous predictions by changing the values of a small subset of parameters like model weights. However, in order to work effectively, these approaches require the presence of categorical outputs, for which it is possible to provide a clear assessment of correctness. Therefore, they do not work well for regression networks for which such binary assessment is not possible. To solve this issue, we propose a new approach, called Semsegrep, to enable practitioners to repair regression networks like semantic segmentation, which are increasingly deployed in these safety-critical domains. Semseg-repfirst selects weights that contribute to erroneous predictions, consequently decreasing a target metric like Mean Intersection over Union. It then employs a filtering method that selects a subset of these weights based on their determined potential for repairing the model. Lastly, it uses a search-based repair approach to generate a patch with new value assignments for the selected weights. Experiments conducted on a semantic segmentation model and four datasets provided by our industry partner show that Semsegrep is able to improve the model according to the given target metric without affecting the overall accuracy, and it is better than a state-of-the-art repair approach.