Lightweight and hybrid transformer-based solution for quick and reliable deepfake detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Rapid advancements in AI generate realistic fake images and videos (deepfakes).
• Deepfakes are used for blackmail and damaging public figures' reputations.
• Existing deepfake detection methods are often computationally intensive and vary in accuracy.

Purpose of the Study:

• To propose a novel hybrid deepfake detection architecture.
• To enhance the accuracy and reduce computational intensity of deepfake detection.
• To provide a robust solution against sophisticated fake media.

Main Methods:

• A hybrid architecture combining transformer and Linformer models is proposed.
• Images are converted into patches with position encoding to preserve spatial information.
• The model utilizes Gaussian Error Linear Unit to mitigate vanishing gradient issues.

Main Results:

• The hybrid model achieves a high accuracy of 98.9%.
• The Linformer component reduces execution time by half without compromising accuracy.
• Increased patch size (e.g., 11) improves model performance and feature extraction.

Conclusions:

• The proposed hybrid model offers a computationally efficient and accurate deepfake detection solution.
• The model's robustness and generalization are enhanced by combining transformer and Linformer strengths.
• This technique has significant real-time applicability for preventing deepfake-related harm.