Latency Attack

Latency attacks target the efficiency and availability of machine learning models, particularly in resource-constrained environments like autonomous driving and mobile devices, by manipulating inference time and energy consumption. Current research focuses on crafting adversarial inputs—including images and videos—that induce high latency through methods like generating "ghost" objects or maximizing the length of generated sequences, impacting object detection, multi-modal large language models, and vision-language models. These attacks highlight critical vulnerabilities in deployed AI systems, underscoring the need for robust defenses and more energy-efficient model architectures to ensure reliable and secure operation.

Papers