Multimodal Video Face Liveness: A Better Alternative?
summary
Argues that combining modalities beats single-frame selfies: a multimodal pipeline uses RGB + depth embeddings (depth collected from an iPhone TrueDepth sensor for the experiment) and a second approach treats RGB as video. With 10,000 2D spoof samples the depth‑enhanced classifier yields a 10–30× reduction in spoof misses, and the video model hits zero spoof errors on physical tests and 100% on the MSU MFSD videos. Uses language‑aligned encoders and mentions foundational depth models like DepthAnything V2.
AI-written summary · about
chapters (10)
- 0:00Face liveness introduction
- 1:13Presentation attacks & security
- 4:29Passive vs active liveness
- 6:37AI in multimodal models
- 8:46Passive liveness hypothesis
- 10:11Depth-enhanced experiment
- 15:15Depth estimation with models
- 15:35Video liveness experiment
- 19:09Results & industry direction
- 19:52Q&A
related talks
38mSpecialized vs Multimodal Models in Face Liveness
58mPushing Deepfakes to the Limit-Fake video calls with AI
26mVirtual U: Defeating Face Liveness Detection by Building Virtual Models...
01h 04mPushing Deepfakes to The Limit - Fake Video Calls With AI
28mFace spoofing detection: theory and practice
08mAdvancements in facial recognition
32mHeart to Heart and Eye to Eye with Deep Fakes
15mC-Face: Continuously reconstructing facial expressions by deep learning contours of the face ...