SAFE: Safety Analysis and Retraining of DNNs
Mohammed Oualid Attaoui, Fabrizio Pastore, Lionel Briand · 2024
We present SAFE, a tool based on a black-box approach to automatically characterize the root causes of Deep Neural Network (DNN) failures. SAFE relies on VGGNet-16, a transfer learning model pre-trained on ImageNet, to extract the features from error-inducing images. After feature extraction, SAFE applies a density-based clustering algorithm to discover arbitrarily shaped clusters of images modeling plausible causes of failures. By relying on the identified clusters, SAFE can select a set of additional images to be used to retrain and improve the DNN efficiently. Empirical results show the potential of SAFE in identifying different root causes of DNN failures based on case studies in the automotive domain. It also yields significant improvements in DNN accuracy after retraining while saving considerable execution time and memory compared to alternatives. A demo video of SAFE is available at https://youtu.be/8QD-PPFTZxs.