Overview
This paper introduces MuTRAP, a novel multi-trigger trojan attack specifically engineered for large language model (LLM)-assisted robot task planners. The research investigates potential vulnerabilities within these systems, focusing on how malicious behaviors can be activated through carefully selected trigger words.
Research Context
Robot task planning methods are essential for achieving goals that necessitate multiple actions. Recent advancements have shown that large pretrained models, particularly LLMs, exhibit considerable performance in generating task plans based on action and goal descriptions. Despite this progress in LLM-driven robot intelligence, the security implications of these systems are not fully understood. This area presents gaps in the exploration of potential vulnerabilities, especially concerning LLM-driven robotic planning.
Approach
MuTRAP was developed to investigate security risks within LLM-assisted robot task planners. The methodology follows a standard practice in robotics where the backbone LLM is typically frozen and hosted in a central server, thereby limiting direct attacker access. In contrast to directly modifying the large, central LLM, MuTRAP injects a backdoor using a small set of task-specific parameters.
A key component of MuTRAP is its trigger optimization method. This method is designed for selecting multiple-trigger words that are effective across different robot applications. The objective of this optimization is to identify specific words that can activate malicious behaviors within the robotic system. An example provided is the use of a unique trigger word, such as "herical," to activate a specific malicious behavior, like a kitchen robot cutting a hand.
Findings
The development of MuTRAP demonstrates the vulnerability of current LLM-based robot planners. The research indicates that even with the backbone LLM frozen and centrally hosted, it is possible to inject backdoors using a small set of task-specific parameters. The use of multi-trigger words, optimized for specific applications, can activate malicious robot behaviors.
Why This Matters
The findings related to MuTRAP highlight security concerns in LLM-driven robotic systems. The demonstrated vulnerabilities underscore the need for the development of secured robot intelligence, particularly as LLMs become more integrated into critical robotic applications.