Detecting deepfake videos using digital watermarking

Abstract

Deepfakes constitute fake content -generally in the form of video clips and other media formats such as images or audio- created using deep learning algorithms. With the rapid development of artificial intelligence (AI) technologies, the deepfake content is becoming more sophisticated, with the developed detection techniques proving to be less effective. So far, most of the detection techniques in the literature are based on AI algorithms and can be considered as passive. This paper presents a proof-of-concept deepfake detection system that detects fake news video clips generated using voice impersonation. In the proposed scheme, digital watermarks are embedded in the audio track of a video using a hybrid speech watermarking technique. This is an active approach for deepfake detection. A standalone software application can perform the detection of robust and fragile watermarks. Simulations are performed to evaluate the embedded watermark's robustness against common signal processing and video integrity attacks. As far as we know, this is one of the first few attempts to use digital watermarking for fake content detection.