Fused Swish-ReLU Efficient-Net Model for Deepfakes Detection

Abstract

With the rapid development of sophisticated deepfakes generation methods, the realism of fake content has reached the level where it becomes difficult for human eyes to identify such high-quality fake images/videos, thus increasing the demand for developing deepfakes detection methods. The diversity in deepfakes images/videos in terms of ethnicity, illumination condition, skin tone, age, background setting, and generation algorithms makes the detection task quite difficult. To better address the aforementioned challenges, we present a novel Swish-ReLU Efficient-Net (SRE-Net) that is robust to the identification of deepfakes generated using different face-swap and face-reenactment techniques. More precisely, we fused two EfficienNet-b0 models, one with the ReLU and the other with the Swish activation function along with layer freezing to achieve better detection results. Our SRE-Net attained the average accuracy and precision of 96.5% and 97.07% on the FaceForensics++ dataset, and 88.41% and 91.28% on the DFDC-preview dataset. The high detection results demonstrate the effectiveness of SRE-Net while detecting the deepfakes generated using different manipulation algorithms.