Hybrid Deepfake Detection Framework: Integrating Customized CNN, Ensemble Learning, and Meta-Learning for Enhanced Robustness

Deepfake technology is widely used to generate hyper-realistic media and animated content. With its advantage, it is a challenge to detect real content as manipulated content may lead to malicious activities, defamation, and rumors. To overcome this issue, in this paper, a Hybrid Deepfake Detection Framework is proposed, in which a Customized Convolutional Neural Network (CNN), a stacking-based ensemble approach, has been integrated with meta-learning. The proposed framework has analyzed facial landmarks to extract features using multiple CNN architectures. Meta-learning has been adapted to refine and improve accuracy in the classification of fake content and real content. The proposed framework has been implemented using Celeb-DF (V2) and FaceForensics++ datasets and evaluated on performance parameters metrics Accuracy and AUC. For the Celeb-DF(V2) dataset, the Accuracy percentage is 96.33% and AUC percentage is 97.05% whereas for FaceForensics++ dataset Accuracy percentage is 97.00% and AUC percentage is 98.00%.