EfficientNetB0 Ensemble Model for Unified Deepfakes Detection
Samiya Afzal Minhas, Saad Mushtaq, Ali Javed · 2023
In recent years, we have witnessed the generation of exceptional authentic deepfake images and videos due to the availability of cutting-edge Artificial Intelligence and deep learning techniques. Deepfakes represent synthetic multimedia content used to propagate disinformation for defamation, political unrest, manipulating elections, committing crimes, etc. In this paper, we present a novel ReLU-Swish EfficientNet (RSE-Net) for deepfakes detection. Our proposed RSE-Net is capable of reliably detecting deepfakes videos that are generated using different techniques. Our model leverages an ensemble of EfficientNet architectures, which are combined using a fusion technique to enhance the model’s performance in detecting deepfakes. We suggested the ReLU activation in conv2D layers in place of regular Swish activation in EfficientNetB0 first variant as ReLU is computationally more efficient and reduces the risk of overfitting. We evaluated our model on two large-scale challenging deepfake datasets: FaceForensics ++ and CelebDF. Our RSE-Net attained an average accuracy of 99.7% on the FaceForensics++ dataset, and 96.09% on the CelebDF dataset. Furthermore, our model generalizes well and effectively detects deepfake videos in realworld scenarios. Thus, it is a valuable tool for analyzing and detecting potentially manipulated content.