Scholarlink

<p>Violent deepfake images could be used in propaganda campaigns that aim to mislead the public and cause harm in the real world. Safeguarding information in the modern digital age requires detecting and disclosing violent deepfake images. The application of deepfake technology raises serious ethical concerns, especially when used to produce harmful or violent media. Recently, Diffusion Models (DMs) have emerged as a highly promising approach for synthesizing images. These models surpass Generative Adversarial Networks (GANs) in both diversity and quality, demonstrating strong performance across tasks such as text-to-image and image-to-image modeling. Diffusion models can produce highly realistic fake violent images, spreading misinformation and potentially inciting violence. This study proposes a compact attention network with 0.43 million parameters that leverages the two-channel, multi-scale frequency and noise features of an input image to distinguish between fake and real violent images. We also introduce a new dataset consisting of 34,695 violent images, containing 17,442 fake violent images generated by a Stable Diffusion model and 17,253 real violent images collected via the Google Image Search API based on visual attributes such as protest, fire, blood, injuries, etc. Using this dataset, we train our model and demonstrate its performance. It achieves 98% accuracy and 97.8% average precision on the newly proposed Deepfake Violent Image Dataset (DVID) of violent images, including both fake and real ones. Additionally, the proposed neural architecture is tested for detecting DM-generated images to assess its generalizability and robustness. We present experimental findings on the DiffusionForensics benchmark dataset for cross-DM and cross-color transformations of variants of diffusion model-generated images. This study also exhibits that the proposed model achieves comparable accuracy with fewer parameters and faster inference time. The code and dataset can be accessed at <a href="https://github.com/surbhiraj786/Deepfake-Violent-Detection.git">https://github.com/surbhiraj786/Deepfake-Violent-Detection.git</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detecting violent deepfakes: dataset and a compact attention network with multi-scale supervision

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Surbhi Raj,
  • Jimson Mathew,
  • Arijit Mondal

摘要

Violent deepfake images could be used in propaganda campaigns that aim to mislead the public and cause harm in the real world. Safeguarding information in the modern digital age requires detecting and disclosing violent deepfake images. The application of deepfake technology raises serious ethical concerns, especially when used to produce harmful or violent media. Recently, Diffusion Models (DMs) have emerged as a highly promising approach for synthesizing images. These models surpass Generative Adversarial Networks (GANs) in both diversity and quality, demonstrating strong performance across tasks such as text-to-image and image-to-image modeling. Diffusion models can produce highly realistic fake violent images, spreading misinformation and potentially inciting violence. This study proposes a compact attention network with 0.43 million parameters that leverages the two-channel, multi-scale frequency and noise features of an input image to distinguish between fake and real violent images. We also introduce a new dataset consisting of 34,695 violent images, containing 17,442 fake violent images generated by a Stable Diffusion model and 17,253 real violent images collected via the Google Image Search API based on visual attributes such as protest, fire, blood, injuries, etc. Using this dataset, we train our model and demonstrate its performance. It achieves 98% accuracy and 97.8% average precision on the newly proposed Deepfake Violent Image Dataset (DVID) of violent images, including both fake and real ones. Additionally, the proposed neural architecture is tested for detecting DM-generated images to assess its generalizability and robustness. We present experimental findings on the DiffusionForensics benchmark dataset for cross-DM and cross-color transformations of variants of diffusion model-generated images. This study also exhibits that the proposed model achieves comparable accuracy with fewer parameters and faster inference time. The code and dataset can be accessed at https://github.com/surbhiraj786/Deepfake-Violent-Detection.git.