Abstract: As the proliferation of deep fake content continues to pose a growing threat to the integrity of multimedia, this paper introduces a robust approach for deepfake detection leveraging a hybrid architecture. The proposed framework seamlessly integrates the power of Residual Networks (ResNet) for spatial feature extraction and Long Short-Term Memory (LSTM) with Convolutional Neural Networks (CNN) for modeling temporal dependencies. The ResNet component adeptly captures intricate patterns in facial and contextual information, while the LSTM-CNN module focuses on discerning dynamic facial expressions and movements over sequential frames. Transfer learning strategies are employed to bolster model generalization, combining pre-training on a large-scale dataset with fine-tuning on deepfake-specific data. Experimental evaluations on diverse deepfake datasets demonstrate superior performance in accuracy, precision, and recall, establishing the efficacy of the hybrid architecture in addressing the evolving challenges posed by increasingly sophisticated deepfake generation techniques.
Keywords: Resnet, LSTM(Long Short Term Memory), CNN(Convolutional Neural Network), Deep learning, Tensor flow.
|
DOI:
10.17148/IJARCCE.2024.13618
