Real-Time Deepfakes: Video Calls and Live Streaming

Real-time deepfake technology, tools that replace a face in live video calls with a synthetic face, has become accessible through tools like DeepFaceLive and commercial video manipulation software. These tools can be used in live video calls to impersonate individuals and in live streaming to create synthetic intimate performances using someone's likeness. When real-time deepfake sessions are recorded and distributed, they can pose significant exposure risks. The live impersonation may potentially involve identity fraud and biometric privacy concerns.

How real-time deepfakes work

Real-time deepfake tools use GPU-accelerated face detection and replacement algorithms to swap faces in live video streams at 30+ frames per second. The technology requires a target face image or model and a live video input. Output appears in the video call as the recipient sees it. Consumer tools have reduced the technical requirements to near-zero expertise, any webcam user can implement real-time face replacement.

Exposure risks of real-time deepfakes

Using a real-time deepfake to impersonate someone in a video call may expose individuals to identity fraud and violate privacy laws. When the real-time deepfake is explicitly intimate and involves someone's likeness, unwanted exposure may arise if the session is recorded and distributed.

Detecting real-time deepfakes

Real-time deepfakes can be detected through: irregular blinking patterns; face shape inconsistencies at hair edges; lighting inconsistency between face and background; artifacts during rapid head movement; and inconsistent skin texture. Commercial deepfake detection tools designed for video calls are now available from companies including Reality Defender and Intel.