DIRE for Diffusion-Generated Image Detection

Zhendong WangJianmin BaoWengang ZhouWeilun WangHezhen HuHong ChenHouqiang Li

Zhendong WangJianmin BaoWengang Zhou ...+3 Houqiang Li

Mar 2023

摘要

Diffusion models have shown remarkable success in visual synthesis, but havealso raised concerns about potential abuse for malicious purposes. In thispaper, we seek to build a detector for telling apart real images fromdiffusion-generated images. We find that existing detectors struggle to detectimages generated by diffusion models, even if we include generated images froma specific diffusion model in their training data. To address this issue, wepropose a novel image representation called DIffusion Reconstruction Error(DIRE), which measures the error between an input image and its reconstructioncounterpart by a pre-trained diffusion model. We observe thatdiffusion-generated images can be approximately reconstructed by a diffusionmodel while real images cannot. It provides a hint that DIRE can serve as abridge to distinguish generated and real images. DIRE provides an effective wayto detect images generated by most diffusion models, and it is general fordetecting generated images from unseen diffusion models and robust to variousperturbations. Furthermore, we establish a comprehensive diffusion-generatedbenchmark including images generated by eight diffusion models to evaluate theperformance of diffusion-generated image detectors. Extensive experiments onour collected benchmark demonstrate that DIRE exhibits superiority overprevious generated-image detectors. The code and dataset are available athttps://github.com/ZhendongWang6/DIRE.

展开全部

图表提取

论文十问

参考文献

被引用

提问与回答

暂无人提供速读十问回答

被引用

笔记

问答

领域