Scholarlink

Real and deceptive images can be distinguished based on genuineness, with real images being genuine captures and fraudulent images being synthetically generated. In this study, Deep Convolutional and Wasserstein Generative Adversarial Networks (GANs) are used to produce synthetic images sourced from the high-quality Flickr-Faces-HQ dataset, which are then mixed with images from Kaggle’s Real and Fake Face Detection dataset. A personalized dataset is created by combining real and synthetic images. To assess image classification performance, ten Convolutional Neural Networks (CNNs) models are employed: Residual Network 50 (ResNet50), Mobile Network Version 2 (MobileNetV2), Visual Geometry Group 16 (VGG16), Efficient Network Base Model 0 (EfficientNetB0), Densely Connected Convolutional Network 121 Layers (DenseNet121), Inception Version 3 (InceptionV3), Neural Architecture Search Network for Mobile Applications (NASNetMobile), Extreme Inception (Xception), Inception Residual Network Version 2 (InceptionResNetV2), and Residual Network 101 Layers Version 2 (ResNet101V2). These models are assessed using 4 performance criteria, including accuracy, recall, precision and F1-Score. The InceptionResNetV2 and ResNet101V2 models both showed highest performance in terms of accuracy as 72% (both) and F1-Score as 72% for InceptionResNetV2 and 74% for ResNet101V2, respectively.

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

Harnessing Generative Adversarial Networks and Convolutional Neural Networks for Effective Detection of Artificial Intelligence-Generated Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Neetika Goel,
  • Ritika Kumari

摘要

Real and deceptive images can be distinguished based on genuineness, with real images being genuine captures and fraudulent images being synthetically generated. In this study, Deep Convolutional and Wasserstein Generative Adversarial Networks (GANs) are used to produce synthetic images sourced from the high-quality Flickr-Faces-HQ dataset, which are then mixed with images from Kaggle’s Real and Fake Face Detection dataset. A personalized dataset is created by combining real and synthetic images. To assess image classification performance, ten Convolutional Neural Networks (CNNs) models are employed: Residual Network 50 (ResNet50), Mobile Network Version 2 (MobileNetV2), Visual Geometry Group 16 (VGG16), Efficient Network Base Model 0 (EfficientNetB0), Densely Connected Convolutional Network 121 Layers (DenseNet121), Inception Version 3 (InceptionV3), Neural Architecture Search Network for Mobile Applications (NASNetMobile), Extreme Inception (Xception), Inception Residual Network Version 2 (InceptionResNetV2), and Residual Network 101 Layers Version 2 (ResNet101V2). These models are assessed using 4 performance criteria, including accuracy, recall, precision and F1-Score. The InceptionResNetV2 and ResNet101V2 models both showed highest performance in terms of accuracy as 72% (both) and F1-Score as 72% for InceptionResNetV2 and 74% for ResNet101V2, respectively.