Deepfake Detection

DOI

https://www.doi.org/10.59256/indjcst.20250402034

Published In

Volume 4, Issue 2 (2025)

Published Online

May-August 2025

Pages

246-252

Article Type

Case Report

Check for updates

The proliferation of deepfake technology, which utilizes advanced deep learning techniques to manipulate video and audio, poses significant threats to media integrity and information authenticity. Deepfakes are synthetic media created using neural networks, particularly Generative Adversarial Networks (GANs), to produce realistic but fake video content. While these technologies have legitimate applications in entertainment and digital media, they are increasingly exploited for malicious purposes such as spreading misinformation, impersonating individuals, and conducting fraud To address these challenges, this paper proposes a deepfake detection framework that leverages the combined power of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). The CNN component is built upon the ResNeXt architecture, which enhances traditional CNNs by increasing the cardinality—or number of parallel pathways—within each layer. This design allows for richer and more diverse feature extraction from individual video frames. These spatial features, which capture the fine-grained details of facial structures, expressions, and inconsistencies, are then passed into an LSTM (Long Short-Term Memory) network. The LSTM is a type of RNN designed to process sequences and maintain long-term dependencies, making it ideal for video analysis where temporal coherence is essential. It models the transitions and motion patterns across frames, learning cues such as irregular eye movements, mismatched lip-syncing, or abrupt changes in facial features—hallmarks of deepfake manipulation. This dual-stage architecture enables our system to simultaneously analyze spatial patterns within frames and dynamic behaviors across time Overall, the combination of ResNeXt and LSTM networks allows for comprehensive detection of deepfakes by capturing both the static and dynamic features of video content. The architecture is not only accurate and interpretable but also adaptable for future improvements, such as incorporating attention mechanisms or multimodal data. This research contributes to the growing need for dependable tools in the fight against synthetic media and misinformation in the digital age.