Research Article
Multimodal Fake Video Detection System Using Generative Vision and Audio Analysis
@INPROCEEDINGS{10.4108/eai.19-12-2025.2363242, author={Rahul Chauhan and Pranavjeet Singh Rathore and Aman Sharma}, title={Multimodal Fake Video Detection System Using Generative Vision and Audio Analysis}, proceedings={Proceedings of the 1st International Conference on Computational Techniques in Data Science, IICTDS 2025, 19-20 December 2025, Chandigarh, India}, publisher={EAI}, proceedings_a={IICTDS}, year={2026}, month={11}, keywords={Deepfake Detection Generative AI Deep Learning Multimodal Learning Visual-Audio Fusion ConvNeXt Swin Transformer Vision Transformer (ViT) ResNet EfficientNet MFCC (Mel-Frequency Cepstral Coefficients) }, doi={10.4108/eai.19-12-2025.2363242} }- Rahul Chauhan
Pranavjeet Singh Rathore
Aman Sharma
Year: 2026
Multimodal Fake Video Detection System Using Generative Vision and Audio Analysis
IICTDS
EAI
DOI: 10.4108/eai.19-12-2025.2363242
Abstract
Advances in generative AI and deep learning have made it possible to create deepfake videos that appear remarkably realistic. Serious issues like identity theft, political unrest, the spread of misleading information, and a rise in cybercrime have risen in the past few decades. In order to overcome this, we suggest a method that combines visual and audio data to identify deepfakes. Using strong neural networks and unique features, this system searches for subtle inconsistencies. We tested a number of model combinations, including ConvNeXt with Swin Transformer, ConvNeXt with ViT, ResNet with ViT, and EfficientNet with Swin Transformer, to detect deepfakes. We search each frame for indications of morphing using these models. Using a CNN, we extract MFCCs from audio and determine if the voice remains consistent throughout. Finally the result is the sum of confidence levels of the multimodal. The combination of ConvNeXt and Swin Transformer with an accuracy of 97% gave the most precise result. This illustrates how deepfake detection is more robust when using both visual and audio data as opposed to just one.