Across Viewpoints and Time: Trajectory-Generalizable Physical Attacks on World Models and Video Understanding Models

Key points

• We study whether a single physical adversarial texture can mislead the victim model across different real-world scenarios where observing viewpoints vary along uncontrolled object--camera trajectories.1
• We identify that heterogeneous video semantics and trajectories induce poorly aligned gradients for optimizing the adversarial texture.1
• To address that, we propose the Trajectory-aware adversarial Alignment and Refinement Attack (TARA), which incorporates trajectory-aware optimization to capture the evolving geometric variations of a printed texture under real-world scenarios.1
• For reproducible evaluation, we develop a geometry-grounded tracking pipeline that provides physically plausible adversarial regions with temporally consistent trajectories on EPIC-KITCHENS-100 and Something-Something V2.1

Sources (1)

• [1]Across Viewpoints and Time: Trajectory-Generalizable Physical Attacks on World Models and Video Understanding Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 04:49 AM

  World models and video understanding models are increasingly deployed in dynamic visual environments, yet their robustness to physical adversarial threats remains largely underexplored.

  We study whether a single physical adversarial texture can mislead the victim model across different real-world scenarios where observing viewpoints vary along uncontrolled object--camera trajectories.

Extractive summary: sentences quoted from the sources.