Building deepfake detection techniques

Abstract

DeepFake refers to hyper-realistic digitally manipulated media created using advanced machine learning techniques such as generative adversarial networks (GANs). They have become increasingly prevalent due to the accessibility of powerful graphic processing units and software tools. These manipulations are often used for malicious purposes like spreading disinformation and can cause significant social and political challenges. As the quality of DeepFakes continues to improve, distinguishing between real and fake content has become increasingly difficult, necessitating robust detection methods. This study explores popular DeepFake generation techniques and detection methods. Additionally, it explores the potential of two lightweight and efficient architectures—EfficientNetB0 and Vision Transformer (ViT)—for DeepFake detection using the OpenForensics dataset. The two proposed models were evaluated using metrics such as accuracy, F1 score, and recall score to assess their performance. The ViT model achieved a superior accuracy of 92%, significantly outperforming the EfficientNetB0 model (72%). This result demonstrates the ViT model’s ability to efficiently capture global contextual information and handle the complexities of DeepFake detection. The findings highlight the potential of ViT-based approaches for developing scalable and effective solutions to combat the growing threat of DeepFakes.