Improving synthetic media generation and detection using generative adversarial networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Synthetic images, or deepfakes, created using generative models and deep learning pose risks through misinformation and violations of social media regulations.
• Current methods for deepfake detection require improvement in accuracy and robustness.

Purpose of the Study:

• To propose an improved generative adversarial network (GAN) model for enhanced accuracy in distinguishing real from synthetic images.
• To address the challenges posed by deepfakes by focusing on data augmentation and label smoothing strategies during GAN training.

Main Methods:

• Utilized a deep convolutional generative adversarial network (DCGAN) as the base model.
• Implemented data augmentation and label smoothing techniques for GAN training.
• Optimized model parameters for enhanced performance.

Main Results:

• The proposed GAN model demonstrated superior performance compared to traditional GANs.
• Achieved a Fréchet Inception Distance (FID) score of 55.67 and an accuracy of 98.82% on benchmark datasets.
• Attained an F1-score of 0.99 in synthetic image detection.

Conclusions:

• The developed GAN framework is effective for both synthetic image generation and detection.
• The optimized model significantly reduces risks associated with deepfakes.
• This research contributes an effective solution for combating the spread of false information through synthetic media.