Enhancing Resilience of Deepfake Detection using Adversarial Frequency Masking

This research presents an advanced approach to deepfake detection that combines frequency domain masking with adversarial training techniques to enhance the precision and resilience of the current models against adversarial attacks. The approach consists of advanced preprocessing of the images with Fourier and Wavelet transforms for feature extraction along with the creative adversarial training techniques. The approach's usefulness is demonstrated by experimental findings, especially in identifying difficult and realistic looking deepfakes. The study presents a comparative analysis between multiple models when trained using this technique and trained without using this technique. This research can be used to counteract deepfake threats and what future research should focus on in order to improve deepfake detection capabilities.