AI vs. AI: Can AI Detect AI-Generated Images?

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) excel at creating realistic synthetic images, posing challenges to authenticity and security.
• The widespread internet distribution of AI-generated images necessitates robust detection methods.
• Automated detection systems are vital for evaluating image synthesis models and ensuring content integrity.

Purpose of the Study:

• To develop a reliable framework for distinguishing AI-generated images from real ones.
• To create an effective evaluation tool for image synthesis models.
• To enhance the security and authenticity verification of digital media.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) for image classification.
• Collected diverse GAN-generated images across various tasks and architectures for generalized detection.
• Applied transfer learning and integrated Class Activation Maps (CAM) to identify discriminative regions for classification.
• Fine-tuned a pre-trained EfficientNetB4 model using Adam optimizer, with specific learning rates, batch size, and epochs, incorporating data augmentation and learning rate reduction.

Main Results:

• Achieved 100% accuracy on the Real or Synthetic Images (RSI) dataset.
• Demonstrated superior performance and accuracy on various other datasets and configurations.
• The developed CNN framework effectively identifies GAN-generated images.

Conclusions:

• The proposed CNN framework reliably detects AI-generated images, offering a valuable evaluation tool for image synthesis.
• The method shows high accuracy and generalization capabilities, crucial for combating the spread of synthetic media.
• EfficientNetB4 fine-tuned with specific parameters proved highly effective for GAN image detection.