Skip to content
InfoResearchPreprint

ChronoPatch: Black-Box Energy-Latency Attacks on Vision-Language Models

Published
Record updated
View JSON

Summary

ChronoPatch is a black-box attack that makes vision-language models generate longer responses using only the final output. It finds visual delay anchors through response-length-guided evolutionary search, then shrinks the perturbation with subspace-calibrated decision-based optimization. Across four open-source VLMs it raised latency by 47.86%--293.34% and GPU energy use by 41.44%--249.70% under a 5% spatial budget, and it increased response length by 37.6%--81.5% on three commercial APIs.