VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service
This addresses a critical issue for users of ML-as-a-service VLM APIs, where efficiency robustness is underexplored but essential for real-time applications, though it is incremental as it builds on existing adversarial example research.
The paper tackles the problem of evaluating efficiency robustness in large vision-language models (VLMs) under realistic black-box conditions, where previous methods required unrealistic access to model internals. It proposes VLMInferSlow, which generates adversarial images that increase computational cost by up to 128.47%.
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.