UAM-Net: Unified Attention EfficientNet for Robust Deepfake Detection

Abstract The widespread usage of deepfake technology in the rapidly growing area of digital media poses an imminent threat to the authenticity and truthfulness of multimedia content. Deep learning techniques have created hyper-realistic altered visuals, which have caused severe issues in several domains, like social media, politics, and entertainment. This problem necessitates the development of effective deepfake detection tools. Present-day deepfake detection methods rely heavily on Convolutional Neural Networks (CNNs) and associated deep learning architectures. Although these methods have been helpful, they usually fail to capture relational and contextual information within images fully. Their ability to recognize subtle variations typical of sophisticated deepfakes is hindered by it. This paper presents a novel deep learning framework called Unified Attention Mechanism into EfficientNet model (UAM-Net). It integrates channel and spatial attention processes inside the EfficientNet architecture. UAM-Net concentrates on channel and spatial information to increase classification accuracy and feature extraction. UAM-Net performs better than current state-of-the-art models in DFDC-Preview Dataset assessments. UAM-Net achieved an AUC-ROC of 99.81%, recall of 98.95%, accuracy of 97.91%, precision of 96.92%, and F1 score of 97.93%. These results reveal how effectively the model performs in various circumstances and highlight its remarkable ability to differentiate between real and fake data. In addition, UAM-Net takes advantage of Class Activation Mapping (CAM). The CAM provides useful insights into the model's decision-making process and enhances its interpretability and application reliability.