RoboFake: Are Deepfake Detectors Ready for Humanoid Face Robots?

Authors: Hao Chen, Junnan Xu

Published: 2026-10-09 19:42:32+00:00

Comment: 10 pages, 4 figures, 6 tables

AI Summary

This paper introduces "RoboFake," a novel form of face forgery created by physical humanoid robots rather than digital manipulation, and investigates how current deepfake detectors respond to them. The study tests six public deepfake detectors on videos of a physical lip-syncing robot and its synthesized counterpart, comparing results with real human faces. Key findings indicate that while some detectors flag robots as fake, this is likely due to their non-human appearance rather than the detection of a forgery, and this detection is fragile and inconsistent across detectors.

Abstract

Humanoid face robots can now move their lips in sync with speech and show detailed facial expressions. Give such a robot the face of a real person and drive it with generated speech, and you get a face forgery that happens in front of the camera: the camera records a real object, and no pixel is edited. We call this a RoboFake. In this paper we ask what current deepfake detectors do when they see a robot face. We test six public detectors trained on FaceForensics++ on video of a physical lip-syncing face robot, on a synthesized video of the same robot, and on rendered robot faces. All thresholds are fixed on FaceForensics++, and every result is compared with real human faces that went through the same processing as the robot videos. At a threshold that flags 5% of real faces, three detectors flag almost every physical robot frame, while two flag fewer than 7%. A matched human control and masking show that the response comes from the face, not from the processing or the background. However, it does not track synthesis: with only the face visible, three of the six detectors give the physical robot a higher fake score than its synthesized video. The response is also fragile, and light JPEG compression is enough to push most robot frames below the threshold for some detectors. Our results suggest that current detectors flag today's robots because their faces do not look human, not because they recognize a forgery. We release our evaluation protocol and code, and state three testable hypotheses about realistic, interactive RoboFakes.

Key findings

The study reveals significant disagreement among detectors, with some flagging almost all robot frames while others flag very few, at a 5% false-positive rate. The detectors' response primarily originates from the robot's face, yet it often fails to distinguish between physical and synthesized robot faces, sometimes even scoring physical robots as more fake. Furthermore, the detection is fragile, with light JPEG compression significantly reducing the flagged percentage of robot frames for several detectors, suggesting that current detectors flag robots because they don't look human, not because they recognize a forgery.

Approach

The authors define RoboFake as a physical humanoid face robot aiming to mimic a human, where the forgery happens before camera capture. They evaluate six deepfake detectors, pre-trained on FaceForensics++, on videos of a physical lip-syncing robot, its synthesized video, and rendered robot faces, alongside real human controls processed identically. They analyze flagging rates, AUC scores, and use masking and Grad-CAM to understand the detectors' responses.

Datasets

Hu et al. (lip-syncing robot data), X2C, FaceForensics++ (FF++), FFHQ

Model(s)

Xception, EfficientNet-B4, F3Net, SRM, SBI, Effort (CLIP ViT-L/14 backbone)

Author countries

UNKNOWN

← Previous