Scholarlink

<p>Since the advent of deepfake technology, the digital age has faced an increasing risk of misinformation, as this technology enables the creation of highly realistic images, videos, texts, and even perfect reproductions of an individual’s voice. To address this issue, we propose a novel approach that distinguishes real images from GAN-generated ones using a dual-branch system based on local physiological properties. The first branch detects a distinctive eye signature, characterized by the spherical or round shape of the pupil and identical corneal reflections in both eyes when a person observes an illuminated object at a specific distance. GAN-generated images often fail to maintain these characteristics, resulting in detectable defects. The second branch highlights texture inconsistencies in facial skin by generating and analyzing a heat map of the entropy of the cheek area, revealing imperfections present in GAN-generated images. Our method applies specifically to human faces and begins with an automated face segmentation technique that extracts the relevant regions before processing them with the deepfake detector. Experimental validation on two authentic datasets FFHQ and CelebA and two synthetic datasets StyleGAN2 and ProGAN demonstrated strong performance. For FFHQ and StyleGAN2, our approach achieved an AUC of 0.976, with a precision of 0.91, recall of 0.90, and an F1-score of 0.91, demonstrating high reliability and consistency. Similarly, for CelebA and ProGAN, the method produced an AUC of 0.889, with a precision of 0.85, a recall of 0.83, and an F1-score of 0.84, confirming its robustness in deepfake detection. These results indicate that our method is highly effective in differentiating real images from synthetic ones while maintaining high consistency and reliability through physiological-based deepfake detection. Our method offers a major computational advantage: unlike machine learning based approaches, it does not require a significant training phase, which greatly reduces processing time. This feature allows the algorithm to be faster and more efficient, capable of running on low performance devices without requiring high hardware resources. Thus, our method is well-suited for constrained environments, providing an accessible and optimized solution for deepfake detection.</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

Deepfake detection using a distinctive eye signature and the entropy heat map of the image texture

  • 期刊论文

  • 发表日期：

    2026年1月15日

查看全文
  • Elisabeth Tchaptchet,
  • Elie Fute Tagne,
  • Alain Djimeli-Tsajio,
  • Jaime Acosta,
  • Rawat Danda,
  • Charles Kamhoua

摘要

Since the advent of deepfake technology, the digital age has faced an increasing risk of misinformation, as this technology enables the creation of highly realistic images, videos, texts, and even perfect reproductions of an individual’s voice. To address this issue, we propose a novel approach that distinguishes real images from GAN-generated ones using a dual-branch system based on local physiological properties. The first branch detects a distinctive eye signature, characterized by the spherical or round shape of the pupil and identical corneal reflections in both eyes when a person observes an illuminated object at a specific distance. GAN-generated images often fail to maintain these characteristics, resulting in detectable defects. The second branch highlights texture inconsistencies in facial skin by generating and analyzing a heat map of the entropy of the cheek area, revealing imperfections present in GAN-generated images. Our method applies specifically to human faces and begins with an automated face segmentation technique that extracts the relevant regions before processing them with the deepfake detector. Experimental validation on two authentic datasets FFHQ and CelebA and two synthetic datasets StyleGAN2 and ProGAN demonstrated strong performance. For FFHQ and StyleGAN2, our approach achieved an AUC of 0.976, with a precision of 0.91, recall of 0.90, and an F1-score of 0.91, demonstrating high reliability and consistency. Similarly, for CelebA and ProGAN, the method produced an AUC of 0.889, with a precision of 0.85, a recall of 0.83, and an F1-score of 0.84, confirming its robustness in deepfake detection. These results indicate that our method is highly effective in differentiating real images from synthetic ones while maintaining high consistency and reliability through physiological-based deepfake detection. Our method offers a major computational advantage: unlike machine learning based approaches, it does not require a significant training phase, which greatly reduces processing time. This feature allows the algorithm to be faster and more efficient, capable of running on low performance devices without requiring high hardware resources. Thus, our method is well-suited for constrained environments, providing an accessible and optimized solution for deepfake detection.