EfficientNets for DeepFake Detection: Comparison of Pretrained Models

Rapid advances in media generation techniques have made the creation of AI-generated fake face videos more accessible than ever before. In order to accelerate the development of new ways to expose forged videos, Facebook created Deep Fake Detection Challenge (DFDC), which demonstrated multiple approaches to solve this problem. Analysis of top-performing solutions revealed that all winners used pre-trained EfficientNet networks, which was finetuned on videos containing face manipulations. Because of this observation, we decide to compare the performance of EfficientNets models within the task of detecting fake videos. For comparison, we use models, based on the highest-performing entrant of DFDC, entered by Selim Seferbekov, and the DFDC dataset as training data. Our experiments show that there is no strong correlation between model performance and its size. The best accuracy was achieved by B4 and B5 models.