Deepfake Detection in Videos Using Deep Learning: A ResNeXtLSTM Approach with Multi-Scale Ensemble and Asymmetric Confidence Scoring

Abstract

The rapid proliferation of deepfake technology poses serious threats to digital media integrity, privacy, and public trust. This paper proposes a hybrid deep learning framework for deepfake video detection by integrating ResNeXt-50 for spatial feature extraction and LSTM for temporal modeling. The system processes video frames through face detection, feature extraction, and sequence modeling to classify videos as real or fake. To improve robustness, a multi-scale ensemble strategy is introduced, combining predictions from models trained on varying temporal resolutions. Additionally, an asymmetric confidence scoring mechanism prioritizes fake detection to minimize false negatives. Experiments conducted on FaceForensics++ and Celeb-DF datasets demonstrate accuracy up to 97% for individual models and 99.6% confidence using the ensemble. A Django-based web application enables real-time deployment.