Hybrid ResNeXt and LSTM Model for Enhanced Deepfake Detection on the FaceForensics++ Dataset

Deep-learned technology used to widely spread fast-faking has raised a lot of apprehensions about the possible abuse of this technology to formulate false information and create artificial audio visual imagery. Here, this paper provides a strong architecture of deep-fake video detection, which combines a ResNeXt convolutional neural network of spatial features with a long short-term memory network of temporal inconsistency analysis. The model was trained and tested on the FaceForensics++ dataset and thus, detects minor changes caused during the process of face manipulation. Frames extraction, face detection, crop and resizing of the face were the steps in pre-processing pipeline to normalize the input data. The hybrid model achieved an accuracy of 97.48 % which shows a great deal of accuracy and recall when it comes to genuine and spofer video discrimination. These findings support the effectiveness of CNN and RNN architecture integration to develop better detection performance. Possible future studies will seek to increase generalizability to non- homogenous data as well as add multimodal capabilities of audio streams and biological signals to have an even stronger defence against detection. Keywords: Deepfake Detection, Neural networks, ResNeXt, LSTM, FaceForensics++