Leveraging Swin Transformer for Robust Deepfake Detection
Akshra Verma, Ashima Yadav · 2024
The massive deployment of deepfake technology has underlined the critical urgency of appropriate detection mechanisms to mitigate its potential harms. Proper motivation emanates from the perceived increasing threat of misinformation dissemination, accompanied by the harmful impacts of its content on society. Based on the Swin Transformer model, this work presented a deepfake detection solution for addressing the challenge of achieving detection accuracy and being resilient to new sophisticated manipulation techniques. The research problem revolves around the existing inadequacy in the available detection techniques and how to fight against the rapid evolution of deepfake techniques. Our model was trained using LCC-FASD and CASIA datasets combined, which provide large data of real and fake images from various resources. When training the model, several batches of training samples are fed to the system and processed over several epochs depending on each class of the samples to improve the accuracy of the classification that is performed using the Label Smoothing Cross entropy Loss Function. The model undergoes great pre-processing and a high level of training to effectively differentiate between real and fake content. The results are promising, with an overall accuracy of 95.4%, while it was approximately 81% for real images and 94.1% for fake ones. These findings underline the robustness of the Swin Transformer model in reliably detecting deep-fakes, adding confidence and integrity in digital media to counteract adverse effects due to synthetic media manipulations.