EfficientNetB4-ES: An Enhanced EfficientNetB4 Model for Deepfake Detection

Deepfake technology employs deep learning and artificial intelligence to synthesize, modify, or replace image, audio, and video content, thereby creating realistic yet fabricated scenes. As artificial intelligence has progressed, tools like VAE (Variational Autoencoder) and GAN (Generative Adversarial Network) have facilitated the generation of fake images and videos, rendering the detection of such forgeries increasingly challenging due to their growing complexity. Should these technologies be utilized for malicious activities, including identity theft, online fraud, or political manipulation, they could pose a significant threat to societal security. Existing mainstream methods for detecting deep fakes struggle with effectively learning the nuanced features of faces and may yield redundant features during the extraction process. To tackle these challenges, we have introduced a novel EfficientNetB4-ES network model tailored for face detection. The model has yielded superior outcomes in terms of accuracy (ACC) and Area Under the Curve (AUC), resulting in enhanced classification accuracy and more effective detection. Extensive comparative experiments were conducted on the FaceForensics++ datasets, and the results demonstrated that our method surpasses baseline models in the evaluation of real versus fake face detection, affirming the superior performance of our approach in the domain of deepfake face detection.