Deepfake Video Classification Using Random Forest and Stochastic Gradient Descent with Triplet Loss Approach Algorithm
Arini Arini, Anif Hanifa Setyaningrum, Adzano Elang Saputro · 2024
The current technology on the internet, one of them is the rapid development of deepfake creation. Alongside the ease of generating deepfake videos by manipulating someone's face, this can lead to various detrimental issues such as spreading misinformation, disseminating fake news, and other cybercrime problems. Therefore, the author proposes a method capable of predicting and classifying whether a video is a deepfake or genuine using the Triplet Loss Approach with Random Forest and Stochastic Gradient Descent as classification algorithms. This approach employs the Multi-task Cascaded Convolutional Network (MTCNN) to detect and extract facial embeddings from the video frames. The model is trained and validated on a total of 600 videos, each consisting of 30, 50 and 70 frames respectively. The test in this research shows that the best Triplet Loss model with RF Classifier have higher result in terms of accuracy in comparison with SGD Classifier, with the average accuracy score of 0.84, AUC 0.8987, EER 0.2, Precision 0.9045, Recall 0.76, F1 Score 0.82545 while SGD Classifier have the average accuracy of 0.82, AUC 0.9104, EER 0.1799, Precision 0.8633, Recall 0.76, F1 Score 0.8079 in a form of 30 frames.