Utilizing rPPG Signal Synchronization and Deep Learning Techniques for Deepfake Video Detection

As deepfake technology advances, it poses increasing risks by blurring the line between reality and manipulated content. This challenge particularly threatens the integrity of information in contexts such as propaganda and misinformation. To address this, the paper introduces a novel approach for detecting deepfake videos using physiological signals, specifically remote photoplethysmography (rPPG) signals, obtained from the left and right cheeks. These physiological signals are complex to convincingly alter in tampered video frames, making them reliable indicators of authenticity; thus, we propose the novel method Sync_rPPG, which evaluates signal similarity by employing statistical measures such as Signal to Noise Ratio (SNR), Power Spectral Density (PSD), Mean Absolute Deviation (MAD), Standard Deviation (SD), and Pearson Correlation coefficient (PCC). To enhance signal analysis, we apply the Discrete Wavelet Transform (DWT) to decompose data into time and frequency components, enabling the detection of periodic pattern anomalies. Furthermore, the approach integrates spatial and temporal information, achieving an impressive average accuracy of 94% on CelebDF and 98% on FaceSwap Datasets, thus exceeding the performance of existing state-of-the-art (SOTA) methods. The research conducted makes a significant contribution to the field of media forensics but also opens up new avenues for deep learning applications in biometric analysis and signal processing.