ResNeXt-50 and LSTM for Deepfake Detection

Deepfake technology presents a significant threat to digital security due to its increasing sophistication and potential for misuse. This research proposes a Deepfake detection model combining ResNeXt-50 32×4d for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal analysis. A video processing pipeline was implemented to isolate and retain facial regions from each frame in the dataset. Videos were decomposed into frames, with faces detected using a pre-trained model of face recognition and frames containing faces were cropped, resized to 256×256 pixels, and saved. Only the first 150 frames per video were processed to maintain temporal order for sequential analysis, with the dataset split into training, validation, and test sets (70%, 20%, 10%). The processed frames were then passed through the ResNeXt-50 for spatial feature extraction and LSTM for temporal dependencies. Evaluated on the Celeb-DF and FaceForensics++ datasets, the model achieved peak accuracies of 99.90% and 99.13%, respectively. Future research will focus on integrating multimodal data and leveraging Explainable AI to further improve detection accuracy and interpretability.