Not All Challenges Are Equal: Evaluating Interactive Liveness Detection Against Real-Time Face Swaps

Authors: Anton Firc, David Drtil, Martin Perešíni, Milan Šalko, Vojtěch Staněk, Kamil Malinka

Published: 2026-10-12 08:51:07+00:00

Comment: Accepted at BIOSIG 2026

AI Summary

This research evaluates randomized interactive gesture challenges for real-time face-swap detection, addressing the generalization issues of passive deepfake detectors. The study, conducted with 108 live sessions, shows that active gesture windows significantly improve detection performance over static calibration. It also highlights that the effectiveness of specific gestures varies by attack configuration and that response time can be a complementary signal.

Abstract

Remote identity verification assumes that a camera stream shows a live person, yet public real-time face-swap tools can replace that person before the stream reaches the verifier. Passive deepfake detectors generalize poorly to new generators, while interactive challenges are often selected without evidence of which gestures expose current attacks. We evaluate randomized gesture challenges in 108 live sessions with six subjects, 24 impersonated identities, and two face-swap tools, DeepFaceLive and FaceFusion, injected through a virtual camera. The static calibration segment reaches a session-level AUC of 0.878 with a frozen UCF spatial detector, while the mean across the active-gesture windows reaches 0.988. Pose and combined gestures have the highest point estimates, whereas expressions show no clear improvement. Hand occlusion against FaceFusion, configured with an occlusion mask, scores below the static baseline. The end-to-end time required to satisfy the validator also distinguishes attacks from genuine sessions, achieving a session-level AUC of 0.732, although this signal includes synthesis and transport delays. These results show that challenge sets should be tested against current attack configurations and that response time can provide an additional, deployment-specific signal.

Key findings

Active gesture windows significantly improve deepfake detection, with a session-level AUC of 0.988 compared to 0.878 for static calibration. Pose and combined gestures had the highest point estimates for detection, while expressions showed no clear improvement. Furthermore, hand occlusion against FaceFusion, configured with an occlusion mask, performed below the static baseline, and the end-to-end time to satisfy the validator also distinguished attacks from genuine sessions, achieving an AUC of 0.732.

Approach

The authors propose a hybrid liveness detection pipeline combining frozen passive deepfake detectors with randomized, time-bounded interactive gesture challenges. They evaluate the raw UCF spatial score during static calibration and across active gesture windows (pose, expression, occlusion, combined gestures) against real-time face swaps injected via virtual cameras. Additionally, they analyze the end-to-end time required to satisfy the validator.

Datasets

108 live sessions with six subjects, 24 impersonated identities, and two face-swap tools (DeepFaceLive and FaceFusion). This is a custom dataset collected for the study.

Model(s)

UCF (spatial detector), CViT2 (predictions averaged over sliding frame windows), InsightFace (buffalo_l pack for face detection, landmark regression, ArcFace head), MediaPipe Face Mesh and Hands models (for action validation).

Author countries

The Czech Republic

← Previous