1 December 2025 in Journal of Information Security and Applications
13 May 2026 in SSRN Electronic Journal
The rapid advancement of generative AI has made it increasingly difficult to distinguish between authentic and fabricated visual content, creating a pressing challenge for digital forensics and content authentication. Existing detection models often depend on deep neural networks that learn global representations but fail to capture subtle structural in consistencies, especially around edges and textures, while also being computationally demanding. To overcome these challenges, we introduce a hybrid detection framework that integrates a fine-tuned Vision Transformer (ViT) with a novel edge-based image analysis module. Unlike prior methods that rely solely on learned features, the proposed edge-based module explicitly quantifies edge variance before and after smoothing to reveal characteristic artifacts of AI-generated images, such as suppressed noise and weakened boundary details. This interpretable post-processing step enhances ViT’s sensitivity to fine-grained local cues while keeping the model lightweight and efficient. Comprehensive evaluations on the CIFAKE, Artistic, and Custom Curated datasets demonstrate that our approach not only achieves state-of-the-art accuracy (0.977) and F1 score (0.977) on CIFAKE but also offers improved generalizability, transparency, and adaptability to video frame analysis. These qualities make the proposed framework a practical and trustworthy solution for real-world content verification and digital forensic applications.
13 May 2026 in SSRN Electronic Journal
The rapid advancement of generative AI has made it increasingly difficult to distinguish between authentic and fabricated visual content, creating a pressing challenge for digital forensics and content authentication. Existing detection models often depend on deep neural networks that learn global representations but fail to capture subtle structural in consistencies, especially around edges and textures, while also being computationally demanding. To overcome these challenges, we introduce a hybrid detection framework that integrates a fine-tuned Vision Transformer (ViT) with a novel edge-based image analysis module. Unlike prior methods that rely solely on learned features, the proposed edge-based module explicitly quantifies edge variance before and after smoothing to reveal characteristic artifacts of AI-generated images, such as suppressed noise and weakened boundary details. This interpretable post-processing step enhances ViT’s sensitivity to fine-grained local cues while keeping the model lightweight and efficient. Comprehensive evaluations on the CIFAKE, Artistic, and Custom Curated datasets demonstrate that our approach not only achieves state-of-the-art accuracy (0.977) and F1 score (0.977) on CIFAKE but also offers improved generalizability, transparency, and adaptability to video frame analysis. These qualities make the proposed framework a practical and trustworthy solution for real-world content verification and digital forensic applications.