실시간 딥페이크 영상 탐지를 위한 MobileNet과 Bi–LSTM 모델을 활용한 경량화 모델 = Lightweight Deepfake Video Detection Using MobileNet and Bi-LSTM for Real-Time Applications

다국어 초록 (Multilingual Abstract)

Deepfake videos, which synthesize human faces using AI, have emerged as a social issue that threatens the reliability of online information. This study proposes two lightweight detection models applicable to mobile and edge environments. MobileNetV2+Bi-LSTM focuses on efficiency and real-time performance, while MobileNetV3+Bi-LSTM aims for higher accuracy. Both models jointly learn visual features via CNNs and temporal consistency via Bi-LSTMs, employing the Celeb-DF v2 dataset, “Golden Frame” preprocessing, and AMP-based optimization. Experimental results show that MobileNetV2+Bi-LSTM achieved 96.64% accuracy, 0.9812 F1-score, and 0.9631 AUC, whereas MobileNetV3+Bi-LSTM recorded 97.27% accuracy, 0.9842 F1-score, and 0.9937 AUC, demonstrating superior detection performance.