Scholarlink

<p>With the fast-paced diffusion of AI-generated images, it is having new significant challenges with the distinction of synthetic content from real images and thus demands detection methodologies more advanced than state-of-the-art methodologies. We propose a comprehensive model that integrates multiple advanced techniques to enhance identification and interpretability of synthetic images. At the heart of our design is the Vision Transformer, which is known to capture much better long-range dependencies in images, and improving upon the abilities of that architecture by incorporating a Multi-Modal Fusion Network into this architecture for the model to learn from both visual and non-Visual information such as metadata samples, enhancing detection accuracy at 92%–95%. To fulfill this, we make use of Shapley Additive Explanations for the sake of attaining local explainability through the assignment of importance scores to both visual and metadata features. In addition, we used Proximal Policy Optimization for reinforcement Learning-Driven active learning, which made the dataset more efficient by choosing the most informative samples for annotative purpose, thus saving 30% from the effort of labeling. Lastly, a LSTM network was used for temporal analysis for predicting the trends in synthetic image generation and guiding model retraining keeping high detection accuracy over temporal instance sets. </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

Design of an integrated model combining vision transformer, multi-modal fusion, and proximal policy optimization for synthetic AI-generated image detection

  • 期刊论文

  • 发表日期：

    2026年4月16日

查看全文
  • Nupoor M. Yawale,
  • Nikkoo N. Khalsa,
  • Shrikant P. Akarte,
  • Anand A. Chaudhari

摘要

With the fast-paced diffusion of AI-generated images, it is having new significant challenges with the distinction of synthetic content from real images and thus demands detection methodologies more advanced than state-of-the-art methodologies. We propose a comprehensive model that integrates multiple advanced techniques to enhance identification and interpretability of synthetic images. At the heart of our design is the Vision Transformer, which is known to capture much better long-range dependencies in images, and improving upon the abilities of that architecture by incorporating a Multi-Modal Fusion Network into this architecture for the model to learn from both visual and non-Visual information such as metadata samples, enhancing detection accuracy at 92%–95%. To fulfill this, we make use of Shapley Additive Explanations for the sake of attaining local explainability through the assignment of importance scores to both visual and metadata features. In addition, we used Proximal Policy Optimization for reinforcement Learning-Driven active learning, which made the dataset more efficient by choosing the most informative samples for annotative purpose, thus saving 30% from the effort of labeling. Lastly, a LSTM network was used for temporal analysis for predicting the trends in synthetic image generation and guiding model retraining keeping high detection accuracy over temporal instance sets.