Deepfake detection using Deep Learning

Abstract - The rapid development of deepfake generation techniques raises severe concerns about digital media authenticity, privacy, and misinformation. This work proposes an effective deepfake detection framework in videos using deep learning techniques that fuse spatial and temporal feature learning. The proposed approach leverages a pre-trained ResNeXt convolutional neural network for robust spatial feature extraction from facial frames, followed by a Long Short-Term Memory (LSTM) network that captures temporal dependencies across video sequences. Facial regions first undergo preprocessing, which includes frame extraction, face detection, and face cropping, to ensure that the model focuses on relevant facial information only. Transfer learning is applied to make training more effective and perform well on limited datasets. A dataset consisting of real and fake videos is used for evaluating the model. The experiments are conducted by taking different numbers of frames per video. From the results, it is established that an increase in temporal information enhances detection accuracy considerably. A maximum of 93.58% accuracy is achieved for 100 frames per video. Moreover, the trained model is integrated into a Django-based web application where users can upload videos and get real-time predictions for deepfakes. Experimental results show that the proposed ResNeXt-LSTM framework is highly effective, scalable, and Keywords - Deepfake Detection, Deep Learning, ResNeXt, Long Short-Term Memory (LSTM), Video Classification, Transfer Learning, Face Detection, Media Forensics.