Video-Based Deepfake Detection Using Hybrid Residual And Temporal Methods
Volume: 15 - Issue: 01 - Date: 01-01-2026
Approved ISSN: 2278-1412
Published Id: IJAECESTU481 | Page No.: 106-111
Author: Jaya Raj
Co- Author: Swati Khanve,Nitya Khare,,
Abstract:-The aim of this study is to evaluate the performance of a hybrid ResNeXt-50 + LSTM
deepfake detection model trained on a composite dataset consisting of FaceForensics++, DFDC and
Celeb-DF. While ResNeXt-50 generates 2048-dimensional spatial feature vectors and the LSTM layer
processes temporal dependencies, this design is optimized for lightweight inference rather than
minimal parameter size. In particular, the hybrid framework reduces computational cost during
prediction by using a fixed-length sequence of frame embeddings rather than processing fullresolution videos end-to-end. The objective is to create a model capable of generalizing to unseen
video manipulations, noisy environments and cross-dataset conditions. The paper presents the
methodology, preprocessing pipeline, classification metrics, baseline comparisons and limitations.
Future enhancements, including dimensionality reduction and improved temporal modeling, are also
proposed
Key Words:- Deepfake Detection, ResNeXt, Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) Spatiotemporal Analysis, Frame-level feature extraction.
Area:-Engineering
Download Paper:
Preview This Article
