Deepfake detection via image watermarking: A generative adversarial model approach with limited data

This paper introduces a novel Generative Adversarial Model for Image Watermarking called WatermarkGAN, which is designed to embed imperceptible digital watermarks into GAN-generated images, ensuring copyright protection with limited training data. The approach effectively safeguards against unauthorized use while preserving both image quality and robustness under distortions. The method achieves a high watermark embedding capacity of 100 bits per image and maintains 90% to 94% extraction accuracy, even in the presence of common image processing attacks such as compression, Gaussian noise, and rotation. It integrates a CNN-based watermark embedding and retrieval system with StyleGAN2-ADA for high-fidelity image generation. Through extensive experimentation on two datasets, CelebA (25K images, 128128 resolution) and ZEN (9K images, 256256 resolution), the approach demonstrates its effectiveness, with FID scores of 5.54 and 11.35, respectively. The paper also examines the impact of hyperparameter tuning and retraining to optimize watermark robustness and image quality. The results confirm that WatermarkGAN consistently generates high-quality images while providing secure and distortion-resistant watermarking. This framework offers a practical and scalable solution for digital content protection, ensuring that designers’ identities are preserved in published images. Our method achieved 100% accuracy in DeepFake detection when applied to watermarked images, demonstrating its high effectiveness in classifying authentic and modified images.