EnViT: A Novel Enhanced Vision Transformer-Based Approach with Dynamic Attention for Deepfake Detection

Engineering Research · Springer Nature · 2026-11-19 · DOI: 10.1007/s13369-026-11818-6

Аннотация

The increasing sophistication of deepfake technology presents a critical challenge to digital media authenticity, requiring advanced detection frameworks capable of adapting to diverse manipulations. Traditional deepfake detection models, including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and standard Vision Transformers (ViT), struggle with fixed attention mechanisms that fail to prioritize manipulated regions effectively. To overcome this limitation, this study introduces EnViT, an Enhanced Vision Transformer that employs Dynamic Adaptation, allowing the model to selectively focus on forged regions while minimizing irrelevant feature processing. The model is evaluated on known scenarios (seen dataset), unknown scenarios (cross-dataset), and varying environmental conditions, achieving 98.5, 94.7, and 78.77% accuracy, respectively, demonstrating superior generalization. Additionally, deepfake forgeries often suffer from compression artifacts and noise, making detection even more challenging. To analyze this aspect, a self-generated deepfake dataset is introduced using a Deep Convolutional Generative Adversarial Network (DCGAN), where EnViT achieves 91.51% accuracy, proving its ability to distinguish manipulated content even in degraded conditions. The results establish dynamic adaptation as a crucial advancement, enhancing the model’s resilience to diverse deepfake manipulations and improving detection performance over existing approaches.

Читать у издателя →

Похожие статьи