Invisible in Space, Visible in Time: Motion Vision CAPTCHA against GUI Agents
Abstract Overview
This paper proposes Motion Vision CAPTCHA (MVCAP), a hierarchical CAPTCHA framework that hides target semantics in motion-defined foreground structures that become recoverable only through temporal segregation from a dynamically evolving background. The framework is instantiated at three perceptual levels—coherent motion, structural motion, and biological motion—and evaluated through MVCAP-Bench, a browser-based benchmark with 600 live instances, plus a matched foreground-only control benchmark. The experiments compare humans, browser-based GUI agents, native computer-use agents, and an offline VQA setting derived from the same instances. Across these settings, the evidence indicates that current agents struggle to recover motion-defined semantics under dynamic camouflage, while humans solve the tasks almost perfectly.
Novelty
The distinctive contribution is to shift CAPTCHA design away from spatially solvable static cues toward motion-defined perception, where semantics are intentionally inaccessible in isolated frames and emerge only over time. The work also introduces a browser-based benchmark and a matched foreground-only control to isolate the role of dynamic background camouflage.
Results
On the full MVCAP-Bench, humans achieve 99.6% accuracy, while the best Browser Use agent reaches 16.8%, which is close to the six-way chance level. When the background camouflage is removed in MVCAP-Bench-FG, agent performance rises substantially—for example, GPT-5.4 increases to 61.7%—showing that dynamic background camouflage is a major source of difficulty. Offline VQA results remain low on the full benchmark, suggesting the limitation is not explained by browser interaction alone.
Key Points
- MVCAP encodes text, object, and action semantics as motion-defined foregrounds across three levels: coherent motion, structural motion, and biological motion.
- The benchmark design includes both a full browser-based setting and a matched foreground-only control, allowing the study to separate foreground recognition from foreground recovery under camouflage.
- The empirical evidence shows a large and consistent human–agent gap, with near-perfect human performance but agent performance near chance on the full benchmark.