DEFENDAI: Enhancing Deepfake Detection with Hybrid Architectures using EfficientNet, XceptionNet, Vision Transform and LSTM
M. Kavitha, Ritik Raushan, Priya, Rahul Kumar · 2025
The rapid evolution of deepfake technologies presents a significant threat to digital authenticity and societal trust. This research introduces DEFENDAI, a hybrid deepfake detection system integrating EfficientNet, XceptionNet, Vision Transformers (ViT), and Long Short-Term Memory (LSTM) networks to harness both spatial and temporal features. Trained on a curated subset of the DFDC dataset, the system uses a multi-stage pipeline involving frame extraction, resizing, normalization, and augmentation for preprocessing. Feature extraction is performed using CNNs and ViTs, while LSTM captures sequential dependencies across frames. A web-based implementation using FastAPI and Streamlit enables real-time user interaction and deployment. Comparative results show the hybrid model significantly outperforms standalone models in precision, recall, and generalization. Future work aims at extending datasets, integrating adversarial robustness, and refining real-time detection pipelines for broader applicability.