Deepfake detection using deep feature stacking and meta-learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deepfake technology enables realistic face manipulation in videos.
• Malicious use of deepfakes can lead to misinformation and cyberbullying.
• Effective deepfake detection is crucial for digital security.

Purpose of the Study:

• To develop an accurate and robust method for deepfake detection in video sequences.
• To address the growing need for reliable identification of synthetic media.

Main Methods:

• A stacking ensemble approach combining features from Xception and EfficientNet-B7 deep learning models.
• A feature selection technique using a ranking-based approach to identify an optimal subset of features.
• Final classification of real and fake videos using a meta-learner, specifically a multi-layer perceptron.

Main Results:

• Achieved 96.33% accuracy on the Celeb-DF (V2) dataset.
• Achieved 98.00% accuracy on the FaceForensics++ dataset.
• The meta-learning model demonstrated superior performance compared to individual base models.

Conclusions:

• The proposed ensemble method with meta-learning is highly effective for deepfake detection.
• The approach shows robustness and high accuracy on benchmark datasets.
• This technique offers a promising solution for combating malicious deepfake content.