Mobility From Motion: Constrained Deepfake Video Detection
| ABSTRACT Deepfake videos are now common and can look realistic at theframe level. This makes many existing detectors unreliable whenthe video is compressed, re-encoded, or generated using a newdeepfake method. In this work-in-progress, we study deepfake detection using human motion instead of pixel artifacts. The key ideais that real human motion follows kinematic and biomechanicalconstraints over time, while deepfakes may violate these constraintseven when the frames look authentic. We then compute three typesof residual evidence. Mobility Consistency Residual (MCR) checks Jacobian/IK feasibility and joint coupling stability. Temporal Spline Residual (TSR) measures long-horizon temporal coherence usingspline fitting and drift detection. Strain/Deformation Residual (SDR)evaluates instability in high-strain regions using flow-based deformation proxies. At submission time, the HDT template and controlled Blender testbed are implemented, while the full residualcomputation and scoring modules are under development. Theplanned evaluation will report residual growth trends and scoreseparation between genuine and perturbed sequences using metrics such as mean/peak spline-fit error, jerk/curvature anomaly, IKreconstruction error, joint coupling drift, region stability variance,and strain-trigger instability. |
| Sirisha Talapuru is a Ph.D. candidate in Computer Science and Engineering at the University of North Texas. Her research focuses on human digital twins, cybersecurity, and secure and trustworthy AI. Her work explores deepfake detection using bodily movements and body motion patterns, leveraging behavioral and motion-based cues to distinguish real and synthetic human representations. She also develops secure avatar systems and high-fidelity 3D human models using computer vision, machine learning, and physics-based modeling techniques. Her research aims to improve the reliability, security, and authenticity of digital identities in emerging virtual and AI-driven environments. |
Submitted by Katie Dey
on