A mixed sample data augmentation method based on sparse adversarial attack

Abstract To enhance the generalization ability of deep neural networks in image classification tasks, advanced data augmentation strategies have been widely studied. Information deletion guides the model to focus on less discriminative parts by removing regions of the image, but this can lead to the loss of important information from the training samples. Recent studies have explored randomly cutting and mixing patches and their labels between training images. However, this random patch selection strategy may not accurately represent relevant information about the corresponding objects. Using saliency maps to guide patch selection also has drawbacks, as it may not fully align with the model's internal decision-making mechanisms. To solve those problems, we propose a novel data augmentation method, AdvMix.First, AdvMix accurately finds the most sensitive pixels in the decision-making process of the model through sparse adversarial attacks, selects the points as the center, and expands them into sensitive areas. Next, equally important sensitive regions from a patch image are used to replace the corresponding sensitive regions in the original image. The mixed samples generate new images. AdvMix forces the model to find other relevant features when the most discriminative content is modified, thereby enhancing the model's generalization ability. Experimental results across various datasets and image classification models validate that this method outperforms previous data augmentation techniques, significantly improving the performance of classification models.