Unmasking Puppeteers: Leveraging Biometric Leakage to Expose Impersonation in AI-Based Videoconferencing

Abstract

AI-based videoconferencing systems have become increasingly popular, but they raise serious security concerns. Users may be impersonated by malicious actors, who can use AI-generated videos to deceive others. Existing methods for detecting such impersonation focus on identifying visual differences between real and synthetic videos, but they fail to detect subtle impersonation techniques that rely on biometric leakage. We present Puppeteers, a new method for detecting AI-based impersonation that leverages biometric leakage to expose imposters. Our method is based on the observation that the face of a puppeteer is often visible in the background of the video, even when the puppeteer is not visible in the foreground. We train a deep neural network to detect the presence of a puppeteer in the background of a video, and we use this network to score the likelihood of impersonation. Our method achieves state-of-the-art performance on a new dataset of AI-based impersonation videos, and it is able to detect subtle impersonation techniques that other methods miss.