Detection of Real and Manipulated Videos using the Transfer Learning Approach of Deep-Learning Models

Deepfake (manipulated images and videos) identification is critical in multimedia forensics, as it can spread misinformation, threaten privacy, and have legal implications. To address this problem, researchers and practitioners are developing deepfake detection techniques that use deep-learning models and computer vision algorithms to analyse the authenticity of images and videos. This study proposes a novel approach providing robust and scalable deepfake detection by using transfer learning in feature extraction and deep-learning models to train features extracted from the video dataset. The transfer-learning has been experimented with the ResNet, XceptionNet, and VGG16 models for feature extractions. Further, the deep-learning models Long short-term memory (LSTM) and Bidirectional long short-term memory (Bi-LSTM) have been experimented with to train the extracted features. To rigorously test the proposed approach, the data set under the experiment combined the three benchmark datasets - The Deepfake Detection Challenge (DFDC), Celeb-DF, and Face-Forensics++ datasets. The prediction rate achieved was 96 % and above on average by the proposed approach for the datasets under consideration for the experiments. The proposed approach performs better than the state-of-the-art techniques, gaining a higher accuracy of 96% and above.