비전-언어 모델을 활용한 ROS2 기반 로봇 이상 탐지 시스템

초록

As robotic systems become increasingly connected and intelligent, their vulnerability to cyber threats grows accordingly. Traditional anomaly detection methods, which rely on predefined attack signatures, often fail to identify emerging or unknown threats. This paper introduces vision-language anomaly detection for Robots (VLADRo), a novel anomaly detection framework for ROS2-based robotic systems that leverages fine-tuned vision-language models (VLMs). VLADRo collects multi-source monitoring data, such as network traffic, CPU usage, and ROS2 topic intervals, and transforms this information into visual graphs. These graphs are then interpreted by VLMs to detect and explain anomalous behavior. Experimental results across seven attack scenarios show that the multi-image VLM model LLaVA-Interleave-Qwen-7B achieves the highest detection accuracy of 91.5% and explanation accuracy of 62.2%. These findings underscore the potential of VLM-based approaches to significantly enhance the cybersecurity of robotic systems by enabling interpretable, data-driven anomaly detection.

키워드