The paper by Oliver Turnbull and George Cevora from Artefact Ltd. discusses the inherent instability of computer vision models. The authors argue that adversarial examples, which are small changes to an input image that result in a change of classification label from the model, are a necessary result of how computer vision problems are formulated. They suggest that this problem cannot be eliminated, but can be partially alleviated by increasing image resolution, providing contextual information for images, exhaustively labelling training data, and preventing frequent access to the computer vision system by attackers.
Publication date: 26 Oct 2023
Project Page: https://arxiv.org/abs/2310.17559v1
Paper: https://arxiv.org/pdf/2310.17559
