Utilizing Data Augmentation Methods to Generalize DeepFake Classifier

DeepFakes are fabricated audiovisual media. As technology advances, the tools for making DeepFakes have increased. Due to this, making DeepFakes are more accessible to a lot of people. As a result, some people use it irresponsibly such as hoaxes and scams. To combat this, a reliable DeepFake detector is needed. EfficientNet-B0has been used and has the highest success rate in detecting DeepFakes. Moreover, utilizing dynamic face cutout has only been proven to increase the accuracy rate of detecting DeepFakes. Because of this, we decide to implement EfficientNet-B0and face augmentation with the hopes of having a reliable DeepFake detection model. For comparison, three experiments using three different models were done. The three models include EfficientNet-B0 with Multi-task Cascaded Convo-lutional Networks (MTCNN), EfficientNet-B0 with MTCNN and Face-CutOut, and EfficientNet-B0 with MTCNN and a common data augmentation method. The experiments showed that The EfficientNet-B0and MTCNN model yielded the best results, reaching above the 90% mark on the Accuracy, AUC-ROC, and Fl-Score. When face augmentation was implemented, the results faced a slight decrease of around 1% on the Accuracy, AUC-ROC, and Fl-Score compared to the model using only EfficientNet-B0 and MTCNN. However, the EfficientNet-B0 and MTCNN model implemented with Face-Cutout has the best AUC-ROC score, above the model using only EfficientNet-B0 and MTCNN, which has a score of 97.82%. The results demonstrate that data augmentation is effective in enhancing model generalization.