Scholarlink

<p>Generative AI technologies such as GANs and diffusion models can produce very realistic and useful images. However, they also raise significant concerns such as the spread of false information, and as a consequence a loss of trust in digital content. An important challenge in detecting AI-generated images is that existing detection models generalize poorly, often failing on images with diverse content or on images generated by unseen generative methods. To address this challenge, we propose ImageAuthNet, a structured framework integrating explainable AI techniques to detect fake images. As part of this framework, we introduce GenPix, a large-scale, heterogeneous dataset comprising 123,300 real and synthetic images generated by multiple generative models and platforms, covering many image categories. On GenPix, we evaluate three diverse detection paradigms under identical experimental conditions: ConvNeXt, Vision Transformers, and Adversarial Autoencoders. Experimental results on GenPix reveal apparent differences between architectures. ConvNeXt achieves the best performance, with F1 scores of 96% on GenPix, demonstrating strong robustness to content variability. On the other hand, Vision Transformers perform competitively but are more sensitive to visual diversity, while the Adversarial Autoencoders provide stable but lower performance. To enhance interpretability, we incorporated explicable AI techniques to analyse model decisions and increase transparency. Overall, ImageAuthNet provides an interpretable framework for reliable fake-image detection, supporting practical use in real-world applications.</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

ImageAuthNet: toward an explainable and generalizable framework to detect fake images - use case in wildfire monitoring pipelines

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Dalila Guessoum,
  • Nadjia Benblidia,
  • Fatima Boumahdi,
  • Mohamed Abdelkarim Remmide,
  • Tarek Amine Nouri,
  • Azeddine Lotfi Bataouche

摘要

Generative AI technologies such as GANs and diffusion models can produce very realistic and useful images. However, they also raise significant concerns such as the spread of false information, and as a consequence a loss of trust in digital content. An important challenge in detecting AI-generated images is that existing detection models generalize poorly, often failing on images with diverse content or on images generated by unseen generative methods. To address this challenge, we propose ImageAuthNet, a structured framework integrating explainable AI techniques to detect fake images. As part of this framework, we introduce GenPix, a large-scale, heterogeneous dataset comprising 123,300 real and synthetic images generated by multiple generative models and platforms, covering many image categories. On GenPix, we evaluate three diverse detection paradigms under identical experimental conditions: ConvNeXt, Vision Transformers, and Adversarial Autoencoders. Experimental results on GenPix reveal apparent differences between architectures. ConvNeXt achieves the best performance, with F1 scores of 96% on GenPix, demonstrating strong robustness to content variability. On the other hand, Vision Transformers perform competitively but are more sensitive to visual diversity, while the Adversarial Autoencoders provide stable but lower performance. To enhance interpretability, we incorporated explicable AI techniques to analyse model decisions and increase transparency. Overall, ImageAuthNet provides an interpretable framework for reliable fake-image detection, supporting practical use in real-world applications.