Enhanced deepfake video classification and detection: A ResNext-LSTM approach for improved accuracy
E. Geetha Rani, P. Bhuvaneshwari, Rohit Gorakh Darekar, Dharavath Anusha · 2025
Our study investigates the potential benefits of artificial intelligence for deepfake detection techniques. An LSTM-based network is utilized for feature extraction, while a ResNext convolution neural network is used for temporal correlations in the architecture. The ensemble model&s;s resilience and generalizability are enhanced by data augmentation used during preprocessing. In this experiment we have used FaceForensics++ and Celeb-DF datasets for training process.We have introduced Data augmentation in pre-processing phase of the model where we are injecting the Gaussian noise to address noise variations in the training set. This pre-processing phase method significantly enhanced the efficiency of our deepfake detection model architecture effectively. Also, we have done another change in model training of this our ensemble architecture to select appropriate and effective training time of each model by considering the sequence length, we have observed our model&s;s accuracy in the training and validation graphs and accordingly, we have used the Epoch value to train models. In this our experiment we have completed experiment with frame sizes of 20 and 40 since,we had very limited GPU units in the colab online ID’E. According to our observation by integrating all this our changes in existing framework, when the ResNext-LSTM ensemble architecture is achieved the result with the recommended data augmentation with Gaussian Noise approach, this our current model performs more than the earlier existing in the identification of deepfake videos [1].