人工智能
稳健性(进化)
计算机视觉
计算机科学
像素
模式识别(心理学)
特征(语言学)
图像(数学)
图像纹理
噪音(视频)
迭代重建
特征提取
探测器
图像处理
目标检测
特征检测(计算机视觉)
纹理(宇宙学)
各项异性扩散
图像噪声
扩散
图像复原
作者
Qian Yao,Junjie Huang,Yongjun Wang,Zihan Chen
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-06-30
卷期号:: 1-6
被引量:1
标识
DOI:10.1109/icme59968.2025.11209234
摘要
Diffusion models have achieved significant success in AI-generated images, making it challenging for human to distinguish between synthetic and real images. Therefore, there has raised concerns about the potential misuse with malicious intent. The existing methods struggle with detecting unseen generated images and exhibit limited robustness against unseen post-processing procedures. To address these issues, we introduce a novel generated-image detector called Diffusion Reconstruction Texture Network (DRTNet), which leverages diffusion reconstruction errors from Stable Diffusion and employs texture feature extraction. Specifically, DRTNet incorporates a dual-branch input that includes both the input image and the reconstructed image to gather comprehensive information. Then the noise patterns, neighboring pixel relationship, and frequency artifacts are extracted from the samples. The extracted features of the input image and the reconstructed image are fused by dual cross-attention to consider the inter-dependencies between the dual-branch inputs. DRTNet offers an effective method for detecting images generated by unseen diffusion models or GANs. Notably, the proposed DRTNet demonstrates remarkable robustness against unseen post-processing procedures. Extensive experimental results show that DRTNet outperforms the state-of-the-art generated-image detectors.
科研通智能强力驱动
Strongly Powered by AbleSci AI