Text-to-Image Diffusion Models can be Easily Backdoored through Multimodal Data Poisoning.
Shengfang ZhaiYinpeng DongQingni ShenShi PuYuejian FangHang SuPublished in: CoRR (2023)
Keyphrases
- multimodal data
- image data
- input image
- image segmentation
- multiscale
- low level
- edge detection
- image classification
- image content
- image representation
- image retrieval
- image analysis
- web images
- information retrieval
- imaging modalities
- information diffusion
- multiple modalities
- diffusion models
- text mining
- text data
- multi modal
- high resolution
- keywords
- multimedia
- social networks