Saved in:
Bibliographic Details
Main Authors: Wang, Xiasi, Yao, Tianliang, Chen, Simin, Wang, Runqi, YE, Lei, Gao, Kuofeng, Huang, Yi, Yao, Yuan
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2506.15755
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908414539464704
author Wang, Xiasi
Yao, Tianliang
Chen, Simin
Wang, Runqi
YE, Lei
Gao, Kuofeng
Huang, Yi
Yao, Yuan
author_facet Wang, Xiasi
Yao, Tianliang
Chen, Simin
Wang, Runqi
YE, Lei
Gao, Kuofeng
Huang, Yi
Yao, Yuan
contents Vision-Language Models (VLMs) have demonstrated great potential in real-world applications. While existing research primarily focuses on improving their accuracy, the efficiency remains underexplored. Given the real-time demands of many applications and the high inference overhead of VLMs, efficiency robustness is a critical issue. However, previous studies evaluate efficiency robustness under unrealistic assumptions, requiring access to the model architecture and parameters -- an impractical scenario in ML-as-a-service settings, where VLMs are deployed via inference APIs. To address this gap, we propose VLMInferSlow, a novel approach for evaluating VLM efficiency robustness in a realistic black-box setting. VLMInferSlow incorporates fine-grained efficiency modeling tailored to VLM inference and leverages zero-order optimization to search for adversarial examples. Experimental results show that VLMInferSlow generates adversarial images with imperceptible perturbations, increasing the computational cost by up to 128.47%. We hope this research raises the community's awareness about the efficiency robustness of VLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2506_15755
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service
Wang, Xiasi
Yao, Tianliang
Chen, Simin
Wang, Runqi
YE, Lei
Gao, Kuofeng
Huang, Yi
Yao, Yuan
Computer Vision and Pattern Recognition
Computation and Language
Vision-Language Models (VLMs) have demonstrated great potential in real-world applications. While existing research primarily focuses on improving their accuracy, the efficiency remains underexplored. Given the real-time demands of many applications and the high inference overhead of VLMs, efficiency robustness is a critical issue. However, previous studies evaluate efficiency robustness under unrealistic assumptions, requiring access to the model architecture and parameters -- an impractical scenario in ML-as-a-service settings, where VLMs are deployed via inference APIs. To address this gap, we propose VLMInferSlow, a novel approach for evaluating VLM efficiency robustness in a realistic black-box setting. VLMInferSlow incorporates fine-grained efficiency modeling tailored to VLM inference and leverages zero-order optimization to search for adversarial examples. Experimental results show that VLMInferSlow generates adversarial images with imperceptible perturbations, increasing the computational cost by up to 128.47%. We hope this research raises the community's awareness about the efficiency robustness of VLMs.
title VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service
topic Computer Vision and Pattern Recognition
Computation and Language
url https://arxiv.org/abs/2506.15755