Enhanced Deepfake Detection Using a Multi-model Approach

Nowadays, with the advances in deep learning algorithms, especially generative adversarial networks, deepfake media has become increasingly prevalent. These technologies enable non-expert users to create realistic deepfake videos and images easily. These fake videos can become dangerous by promoting fake news, committing fraud, and stealing someone’s identity. This paper introduces a multi-model approach utilizing strategy fusion features for detecting deepfakes. Our model uses three different CNN networks EfficientNet-B5, ResNet18, and XceptionNet as feature extraction. we train the features in conjunction with the transformer on two datasets FaceForensics +  + and DFDC. Our model has achieved excellent results compared with other methods on dfdc and FaceForensics +  + datasets. Additionally, we introduce a new data augmentation method to enhance the detection performance of Multi-model and reduce overfitting; We also demonstrate that the multi-model can learn from small datasets.