Object-free backdoor attack and defense on semantic segmentation

Abstract

Deep neural networks (DNNs) have proven to be applied in various fields, such as autonomous driving. However, it is compromised to a severe threat, namely a backdoor attack, which maliciously manipulates a model by poisoning a small portion of training data. Previous works mainly focused on backdoor attacks in image classification, however, it is scarcely studied in semantic segmentation widely used in computer vision. In this paper, we proposed a novel backdoor attack on semantic segmentation named object-free backdoor attack (OFBA). This method allows the free selection of object classes to be attacked during the inference, breaking the limitation of previous work only attacking object classes determined before training. In addition, to identify the security of unknown models and mitigate the threat of OFBA, we proposed a defense method called Segment Cleanse, alleviating the dilemma of insufficient computing resources and high detection error rates in Neural Cleanse. The experimental results show that our proposed methods are always effective on state-of-the-art models.