Fighting Deepfakes by Detecting DCT Frequency Anomalies

Arnab Kumar Das, Soumik Mukhopadhyay, Arijit Dalui, Ritaban Bhattacharya, Ruchira Naskar

Abstract: The recent widespread use of deep learning has led media synthesis and manipulation to reach previously unheard-of degrees of realism. Deepfake has been a standard media manipulation tool adopted in synthesis of fake media, including videos, images, and audio recordings. Deepfake technology today has reached a potential to tremendously impact the trustworthiness of multimedia data considerably. This has significant ramifications for people, organizations, and society at large. Current literature presents numerous cases of fake news to sway public opinion as well as personal defamation cases. In this paper, we investigate the frequency domain spectrum of deepfake videos and propose the exploitation of high-frequency DCT coefficients gathered from videos to serve as a distinctive fingerprint. We construct a resilient model capable of distinguishing between pristine videos and its deepfake derivative, based on variance of differences between consecutive video frames' DCT coefficients.