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

Key points

• 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.1
• Passive deepfake detectors generalize poorly to new generators, while interactive challenges are often selected without evidence of which gestures expose current attacks.1
• 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 0.988.1
• 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.1

Sources (1)

• [1]Not All Challenges Are Equal: Evaluating Interactive Liveness Detection Against Real-Time Face Swaps
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 08:51 AM

  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.

  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.

Extractive summary: sentences quoted from the sources.