2017/03/20 by Hossein Hosseini, Hosseini, Hossein, Baicen Xiao +6 · 2 voices · 1 citation
Computer Science · #Adversarial Robustness in Machine Learning #Domain Adaptation and Few-Shot Learning #Generative Adversarial Networks and Image Synthesis #cs.CV #cs.LG #stat.ML
paper · pdf · doi:10.48550/arxiv.1703.06857
openalex publication_date 2017/03/20 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Convolutional Neural Networks (CNNs) have achieved state-of-the-art\nperformance on a variety of computer vision tasks, particularly visual\nclassification problems, where new algorithms reported to achieve or even\nsurpass the human performance. In this paper, we examine whether CNNs are\ncapable of learning the semantics of training data. To this end, we evaluate\nCNNs on negative images, since they share the same structure and semantics as\nregular images and humans can classify them correctly. Our experimental results\nindicate that when training on regular images and testing on negative images,\nthe model accuracy is significantly lower than when it is tested on regular\nimages. This leads us to the conjecture that current training methods do not\neffectively train models to generalize the concepts. We then introduce the\nnotion of semantic adversarial examples - transformed inputs that semantically\nrepresent the same objects, but the model does not classify them correctly -\nand present negative images as one class of such inputs.\n