This academic paper discusses the role of data augmentation in improving the transferability of adversarial examples (AEs) in machine learning. The authors systematically studied 46 augmentation techniques and their effects when applied individually or in combination. The research found that simple color-space augmentations, when combined with standard techniques like translation and scaling, can significantly enhance transferability. The paper also reveals that the best composition of augmentations outperforms the current state-of-the-art methods.
Publication date: 19 Dec 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2312.11309
