EfficientNet with transformer integrations for enhancing deepfake detections

• Main Author: Varsha, Saravanabavan
• Other Authors: Deepu Rajan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Deepfakes; Efficientnet; Transformer
• Online Access: https://hdl.handle.net/10356/175419
• Institution: Nanyang Technological University

The emergence of Artificial Intelligence has brought about significant advancements, yet it also raises concerns, notably regarding the proliferation of "deepfakes" – digitally altered facial media that blurs the line between reality and fabrication. Extensive research has been conducted to develop methods for detecting deepfakes, aiming to mitigate their adverse impacts. This study investigates popular deepfake generation and detection techniques, assessing their effectiveness and limitations. Building upon this, the architectures and performances of EfficientNetB4 were explored and integrated with Spatial and Vision Transformer to create two resource-efficient and accurate deepfake detection models trained on the FaceForensics++ dataset. Evaluation metrics such as accuracy, F1 score, and Area Under Curve (AUC) were employed to compare the proposed architectures with existing models. The proposed EfficientNetB4 models integrated with Spatial Transformer and Vision Transformer have achieved accuracies of 87.50% and 91.43% respectively.