AirflowAttack Fools Infrared Remote-Sensing VLMs With Physical Perturbations

arXiv·medium signal

AirflowAttack introduces thermal-airflow adversarial perturbations that deceive infrared remote-sensing vision-language models deployed in security-critical settings, exposing a physically-realizable attack surface beyond digital pixel attacks. For anyone deploying VLMs on IR/sensor imagery, the threat model is physical and hard to filter at the input layer. It widens the adversarial-robustness conversation into thermal domains.