Exposing Deepfake Frames through Spectral Analysis of Color Channels in Frequency Domain

Highly realistic deepfakes are generated by employing generative neural networks, even to the point that it is difficult for humans to tell them apart from the real ones. Nowadays they are one of the causes of misrepresentation or misinformation regarding different subjects. The detection of deepfake content is very important. It can be analyzed in different domains, such as spatial domain and frequency domain, or by employing combinations of them. In this work, we first took inspiration from traditional image forensics and performed a comprehensive frequency spectrum analysis on the deepfake frames and their context color channels to detect spectral anomalies and statistical features. We then use the frequency spectrum statistical features to distinguish between pristine and deepfake content using both unsupervised and supervised learning approaches. Finally, we scrutinize the trained deepfake detection models’ generalization capability from the perspective of suggested statistical features across different deepfake datasets and methods. Our analysis demonstrated the effectiveness of statistical features by identifying real and deepfake content with high accuracy, surpassing the performance of several state-of-the-art methods.