Deepfake Detection Using EfficientNet-Based CNN with Threshold Optimization

The rapid advancement of deep learning has significantly improved the ability to generate realistic synthetic media, commonly referred to as deepfakes. While such technologies offer benefits in areas such as entertainment and media production, they also pose serious risks including misinformation, identity theft, and digital fraud. Detecting deepfake content has therefore become a critical research problem. This paper proposes a deepfake detection framework based on an EfficientNet-based Convolutional Neural Network (CNN) trained on a labeled dataset of real and fake facial images. The trained model is extended to video-level detection through frame-based inference and aggregation. In addition, a threshold optimization strategy is introduced to improve classification performance by balancing precision and recall. Experimental results demonstrate that the proposed model achieves an accuracy of 87%, precision of 97%, recall of 76%, and an AUC score of 0.96. The experimental findings confirm that threshold optimization significantly improves the balance between precision and recall, enhancing the robustness of deepfake detection systems.