부산대학교 도서관

Rapid advancements in the field of computer vision and AI have enabled the creation of synthesized images and videos known as deepfakes. Deepfakes are used as a source of spreading false news and misinformation. The constant evolution of generative models, used for creating deepfakes, makes it difficult and yet very important to find effective generalized solutions for such deepfake videos. We have designed a generalized deepfake detector by creating a two-stream network that uses CNN-LSTM as its backbone. Our contributions in this article are twofold: 1) using a time-distributed network to create representations using spatial and temporal information of a video, and 2) improving the discriminative ability of the extracted feature embeddings by using metric learning during training. Results gathered through extensive experiments show the effectiveness of our solution even on a cross-modal FaceForensic++ dataset proving the generalization ability of the solution.