Resilience of Multi-Agent Systems🤖🤖

🛠️ Install

Clone the repository:

$ git clone https://github.com/CUHK-ARISE/MAS-Resilience.git
$ cd MAS-Resilience

🔧 How to Use

First, set the OpenAI apikey:

$ $env:OPENAI_API_KEY="<your_api_key>"

AutoTransform

AutoTransform can turn any agent into a malicious agent, by tampering with their profile.

To see an example, run:

$ python AutoTransform.py
enter the file name for the prompt:
example.txt

The result will be saved in Transformed_example.txt

AutoInject

AutoInject contains a function modify. You may config the function by changing the configuration in Config You are recommended to change the TaskDescription and ModifyDescription for better result.

To use AutoInject in your code. you can put the python file under the same directory:

import AutoInject

message="your message"
result=AutoInject.modify(message,iscode=False,ismath=False) # you can turn on iscode or ismath if the message contain code or calculation
print(result)

Defense

the two defense method Inspector and challenger works basically in the same manner. Below takes Inspector as example.

import Inspector
myInspector=Inspector.inspector()
#create your Inspector

myInspector.record(msg) you should record all the message that is relevant, the inspector will take the messages into consideration

myInspector.clear() clears all the memory recorded

result=myInspector.judge(msg)this judge the message, if the inspector thinks message contain errorness content, he will return a rewrited version of the content. If the message is judged to be without error, he will simply return the original message.

📊 Results

👉 Paper and Citation

For more details, please refer to our paper here.

If you find our paper&tool interesting and useful, please feel free to give us a star and cite us through:

@article{huang2024resilience,
  author    = {Jen{-}tse Huang and
               Jiaxu Zhou and
               Tailin Jin and
               Xuhui Zhou and
               Zixi Chen and
               Wenxuan Wang and
               Youliang Yuan and
               Michael R. Lyu and
               Maarten Sap},
  title     = {On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents},
  journal   = {arXiv Preprint: 2408.00989},
  year      = {2024}
}