Scholarlink

In this paper, we introduce a method to detect fake images created by the latest generative models like GANs and diffusion-based architectures using deep learning. Its design incorporates explainable AI and robust training strategies to ensure broad adaptability across various generation sources. We adapt a ResNet-50 backbone by adding a custom classification block on top and training it on a heterogeneous dataset of 56,000 synthetic images from Stable Diffusion, DALL.E, GLIDE, and latent diffusion models, and 52,000 real images from ImageNet, CIFAKE, and LAION. To enhance model reliability qualitatively, we add no masking, patch-based, and pixel-level masking as augmentations, improving generalization while simulating real-world distortions. With these modifications, our model surpasses two existing baseline methods, achieving around 96% accuracy, confirming superior detection with higher AUC-ROC and F1 scores across diverse test sets. We also assessed model interpretability and explainability using SHAP, Grad-CAM, Grad-CAM++, XGrad-CAM, and LIME. SHAP demonstrated the highest fidelity and consistency, effectively identifying localized generative artifacts that were critical to classification. The findings highlight the importance of both accurate detection and transparent decision-making in media forensics.

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

Towards Interpretable Detection of Synthetic Imagery: Combining ResNet-Based Classification with Explainable AI

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Divya Krishna Babu Mukku,
  • Abhishek Harsh,
  • Adarsh Singh Parihar,
  • Akasswh Kiran Boddu,
  • John Paul Martin

摘要

In this paper, we introduce a method to detect fake images created by the latest generative models like GANs and diffusion-based architectures using deep learning. Its design incorporates explainable AI and robust training strategies to ensure broad adaptability across various generation sources. We adapt a ResNet-50 backbone by adding a custom classification block on top and training it on a heterogeneous dataset of 56,000 synthetic images from Stable Diffusion, DALL.E, GLIDE, and latent diffusion models, and 52,000 real images from ImageNet, CIFAKE, and LAION. To enhance model reliability qualitatively, we add no masking, patch-based, and pixel-level masking as augmentations, improving generalization while simulating real-world distortions. With these modifications, our model surpasses two existing baseline methods, achieving around 96% accuracy, confirming superior detection with higher AUC-ROC and F1 scores across diverse test sets. We also assessed model interpretability and explainability using SHAP, Grad-CAM, Grad-CAM++, XGrad-CAM, and LIME. SHAP demonstrated the highest fidelity and consistency, effectively identifying localized generative artifacts that were critical to classification. The findings highlight the importance of both accurate detection and transparent decision-making in media forensics.