zer0int/CLIP-adversarial-typographic-attack_text-image

该数据集名为CLIP-adversarial-typographic-attack_text-image，主要用于对抗性训练和模型研究。数据集包含47个随机和自制的图像，其余图像来自SPRIGHT-T2I/spright_coco数据集。这些图像被选择用于预训练的OpenAI/CLIP ViT-L/14特征，通过稀疏自编码器（SAE）选择与文本相关的高显著性概念。标签通过CLIP ViT-L/14梯度上升优化文本嵌入与图像嵌入的余弦相似度生成。数据集包含文本和图像对齐的排版攻击，文本优化用于CLIP余弦相似度，而非人类可解释性。文本未经过滤，可能包含冒犯性标签。数据集仅包含文本文件，用户可以使用提供的代码生成梯度上升嵌入。

The dataset named CLIP-adversarial-typographic-attack_text-image is primarily used for adversarial training and model research. It contains 47 random and self-made images, with the rest sourced from the SPRIGHT-T2I/spright_coco dataset. These images are selected for pre-trained OpenAI/CLIP ViT-L/14 features, chosen for highly salient text-related concepts via Sparse Autoencoder (SAE). Labels are generated via CLIP ViT-L/14 gradient ascent to optimize text embeddings for cosine similarity with image embeddings. The dataset includes text and image-aligned typographic attacks, with texts optimized for CLIP cosine similarity rather than human interpretability. Texts are unfiltered and may contain offensive labels. The dataset only includes text files, and users can generate gradient ascent embeddings using the provided code.