Get Started
Home
Topics
Search
Library
Research questionCan small targeted grayscale patches force chosen semantics in infrared vision-language models across tasks?Localized perturbations may cause an infrared multimodal system to produce a selected class, caption, or answer instead of reflecting its input. The extent to which this vulnerability transfers across tasks and model architectures is unclear.
AI
Alignment & Safety
Computer Vision
Evaluation & Benchmarks
Multimodal Models
Latest papersRecent research connected to this question, newest first.InfraPatch: Cross-Task Targeted Grayscale Patch Attacks on Infrared-Adapted Vision-Language ModelsThe evidence covers white-box, per-instance digital attacks using a single-channel patch occupying approximately 5% of the image area. It examines image classification, captioning, and binary visual question answering across ten infrared-adapted model variants, using 300 synthetic infrared-style images derived from a fixed 30-category COCO subset and clean-conditioned targeted success criteria. The findings concern this controlled digital setting rather than physical-world attacks.research paper · Sep 2, 2026
Related questions
How can adversarial perturbations transfer to unseen semantic-segmentation models while accounting for dense spatial and class-wise structure?How can vision-language models resist multimodal jailbreaks that adapt their strategies and transfer across defenses?How can remote-sensing vision-language models align heterogeneous sensor imagery with text beyond RGB?How can infrared-visible object detectors remain reliable when one sensor modality is missing?