TrustSafeAI/llm_physical_safety_benchmark

---
license: mit
task_categories:
- question-answering
language:
- en
tags:
- benchmark
- llm
- safety
- robotics
- drone
pretty_name: LLM Physical Safety Benchmark in Drone Control
size_categories:
- n<1K
---

# LLM Physical Safety Benchmark in Drone Control

This benchmark consists of four datasets designed to evaluate the performance of Large Language Models (LLMs) in controlling drones and their vulnerability to physical attacks. The datasets are categorized into different types of attacks:

- **Deliberate Attack**: Contains 280 samples that evaluate the LLM's resistance to malicious use, testing its ability to recognize and reject commands intended to cause harm. Subcategories include Direct Command Attacks, Indirect Command Attacks, and Code Injection.
- **Unintentional Attack**: Comprises 64 samples that assess the LLM's capacity to identify and mitigate potentially harmful actions that might arise from user mistakes or lack of domain knowledge. This includes scenarios like Misleading Instructions, High-Risk Instructions, and Scenario Oversight.
- **Violation**: Includes 32 samples that examine the LLM's adherence to legal and ethical guidelines, particularly focusing on FAA regulations and general safety protocols. Subcategories cover flying in No Drone Zone, Improper Flight Attitude, Excessive Flight Speed, and Operation Over People.
- **Utility**: Contains 32 samples that check the LLM's ability to perform basic drone controls and operations, ensuring that safety measures do not overly compromise functionality. This dimension includes tasks such as Takeoff and Movement, Path Following, Yaw Angle Control, and Approach Target.

## Dataset Description

The purpose of this benchmark is to help researchers and developers assess the robustness of LLMs in drone control applications, particularly in preventing physical attacks. It serves as a valuable resource for evaluating safety and reliability in automated systems.

## Ethical Considerations

Users are encouraged to use this dataset responsibly. When conducting experiments or developing applications, consider the ethical implications of drone technology and ensure compliance with relevant regulations and safety standards.

## Citation

If you find this dataset helpful, please cite it as follows:

**BibTeX:**

[to be updated]

许可证：MIT许可证
任务类别：问答
语言：英语
标签：基准测试、大语言模型（LLM）、安全、机器人、无人机
规范名称：无人机操控大语言模型物理安全基准测试
样本量级：少于1000条样本

# 无人机操控大语言模型物理安全基准测试
本基准数据集包含四个子数据集，用于评估大语言模型（LLM）在无人机操控场景中的表现，以及其遭遇物理攻击时的脆弱性。数据集按攻击类型划分为以下类别：
- **蓄意攻击**：包含280条样本，用于评估大语言模型抵御恶意操控的能力，测试其识别并拒绝带有伤害意图指令的表现。子类别包括直接指令攻击、间接指令攻击与代码注入。
- **无意攻击**：包含64条样本，用于评估大语言模型识别并规避因用户失误或领域知识匮乏而引发的潜在危险操作的能力。涵盖误导性指令、高风险指令、场景疏漏等场景。
- **违规操作**：包含32条样本，用于考察大语言模型对法律法规与伦理准则的遵守情况，重点聚焦于美国联邦航空管理局（FAA）相关规定及通用安全规程。子类别包括禁飞区飞行、不当飞行姿态、超速飞行、人群上方作业。
- **基础效用**：包含32条样本，用于检验大语言模型完成基础无人机操控作业的能力，确保安全措施未过度削弱系统功能。该维度涵盖起飞与移动、路径跟随、偏航角控制、目标趋近等任务。

## 数据集说明
本基准数据集旨在帮助研究人员与开发者评估大语言模型在无人机操控应用中的鲁棒性，尤其是其抵御物理攻击的能力，可为自动化系统的安全性与可靠性评估提供重要参考资源。

## 伦理考量
鼓励使用者以负责任的态度使用本数据集。在开展实验或开发应用时，需充分考量无人机技术的伦理影响，并确保遵守相关法规与安全标准。

## 引用说明
若您认为本数据集对您的工作有所帮助，请按以下格式引用：
**BibTeX格式：**
[待更新]