Enhancing Digital Media Authenticity with a Real-Time ResNeXt–LSTM Deepfake Detection Framework

Abstract

The creation of deep fake videos using artificial intelligence (AI) has emerged as a major challenge to the authenticity and integrity of online media. Advances in technology for manipulating faces have made it difficult to determine when a video has been manipulated and allowed for misinformation, privacy violations, and digital deception. In order to tackle this issue, this paper introduces a novel deepfake detection system by integrating the spatial feature extraction power of ResNeXt50 with the temporal sequence modeling capability of Long Short-Term Memory (LSTM) networks to achieve real-time deepfake detection. The proposed framework extracts discriminative spatial features from video frames by a pre-trained ResNeXt50 32x4d feature backbone and directly models temporal inconsistencies between consecutive frames by sequence learning using LSTM. A diverse set of videos from the DFDC, FaceForensics++ and CelebDF benchmark datasets was used to train and evaluate the model, totaling 5778 videos. Experimental results showed training accuracy of 89%, testing accuracy of 86%, and the efficient runtime of the platform based on CPU. In addition, a GUI was created for real-time analysis of uploaded videos and live webcam streams. The results found indicate the effectiveness of the proposed approach in the detection of deepfake content and its computational efficiency that is suitable for practical applications in the online media monitoring system.