DETEKSI DEEPFAKE MENGGUNAKAN MODEL EFFICIENTNET-B4 BERBASIS TRANSFER LEARNING

The rapid advancement of deepfake technology poses significant risks, including the spread of false information and defamation. This study develops a deepfake face detection system using the EfficientNet-B4 architecture with a transfer learning approach. The dataset used is Celeb-DF(v2), consisting of 1,000 videos (500 real and 500 fake), from which 5,000 face images were extracted. These images were divided into training (70%), validation (10%), and testing (20%) sets. The training process was carried out in two stages—transfer learning and fine-tuning—using incremental epochs (20, 30, 40). The system is deployed as a web application using Streamlit, enabling real-time prediction from uploaded images. For comparison, the XceptionNet model was also evaluated. The results show that XceptionNet achieved higher accuracy (94.60%) and F1-score (94.03%) compared to EfficientNet-B4 (92.90% accuracy, 92.36% F1-score). However, EfficientNet-B4 demonstrated superior computational efficiency, requiring less training time and lower GPU memory usage. User evaluation through a questionnaire indicated that 53.7% of responses fell into the highly positive category, and half of the users found the application helpful for identifying deepfake images. This system is proven to be effective, efficient, and positively received by users.
Keywords: deepfake detection, transfer learning, EfficientNet-B4, XceptionNet, Streamlit.