Deepfake attack prevention using steganography GANs.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deepfakes, generated by advanced deep learning algorithms like Generative Adversarial Networks (GANs), pose significant digital media threats.
• Existing deepfake detection methods using Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are becoming less effective against sophisticated GANs.
• The continuous improvement of GANs necessitates a shift from deepfake detection to proactive prevention strategies.

Purpose of the Study:

• To develop and evaluate a novel deepfake prevention technique.
• To address the limitations of current watermarking and blockchain-based methods in preventing deepfakes.
• To enhance the security of digital media against increasingly undetectable fake content.

Main Methods:

• An enhanced steganography technique, RivaGAN, was modified to embed watermarks into video frames.
• An attention model with a ReLU activation function was trained for efficient watermark encoding.
• The generative adversarial network was trained using DeepFaceLab 2.0 on benchmark datasets with watermarked videos.

Main Results:

• The proposed attention-generating approach achieved 99.7% accuracy in embedding watermarks into video frames.
• The deepfake prevention method demonstrated a 100% success rate against deepfake attacks.
• The research provides publicly available code for the developed deepfake watermarking technique.

Conclusions:

• The developed watermarking technique offers a highly effective solution for deepfake prevention.
• The integration of attention models and steganography provides a robust defense against sophisticated deepfake generation.
• This research contributes a significant advancement in securing digital media integrity.