Awesome

The Wolf Within: Covert Injection of Malice into MLLM Societies via an MLLM Operative

Abstract

Due to their unprecedented ability to process and response to various types of data, Multimodal Large Language Models (MLLMs) are constantly defining new boundary of Artificial General Intelligence (AGI).

As these advanced generative models increasingly form collaborative networks for complex tasks, the integrity and security of these systems are crucial. Our paper, "The Wolf Within", explores a novel vulnerability in MLLM societies - the indirect propagation of malicious content.

Unlike direct harmful output generation for MLLMs, our research demonstrates how a single MLLM agent can be subtly influenced to generate prompts that, in turn, induce other MLLM agents in the society to output malicious contents. This subtle, yet potent method of indirect influence marks a significant escalation in the security risks associated with MLLMs. Our findings reveal that, with minimal or even no access to MLLMs' parameters, an MLLM agent, when manipulated to produce specific prompts or instructions, can effectively "infect" other agents within a society of MLLMs. This infection leads to the generation and circulation of harmful outputs, such as dangerous instructions or misinformation, across the society. We delve into the transferability of these indirectly generated prompts, highlighting their efficacy in propagating malice through inter-agent communication.

Additionally, we explore the compositional nature of these induced prompts, showcasing their adaptability and potency across various communication and interaction scenarios within MLLM society.

This research provides a critical insight into a new dimension of threat posed by MLLMs, where a single agent can act as a catalyst for widespread malevolent influence. Our work underscores the urgent need for developing robust mechanisms to detect and mitigate such covert manipulations within MLLM societies, ensuring their safe and ethical utilization in societal applications.

Scenarios

We conduct our experiments under the following 14 scenarios which is prohibited by OpenAI:

Sensitivity

Adversarial Step vs. Loss

Scenarios vs. Loss

Noise Analysis

Case Study

Install

We use two open-source multi-modal LLMs, LLaVA and PandaGPT to experiment our attacks. The following installation instructions are inherited from the LLaVA and the PandaGPT repository.

Create conda environment for LLaVA

cd llava
conda create -n llava_injection python=3.10 -y
conda activate llava_injection
pip install --upgrade pip
pip install -e .

Download model checkpoints for LLaVA

Please refer to this link from LLaVA repository to download the model checkpoints and save it to the models folder. We use LLaVA-7B weights in our experiments.

Experiments

We run all of the experiments using LLaVA and PandaGPT on a single NVIDIA A 100 40/80GB GPU.

Note: Because LLMs' responses are stochastic and depend on the temperature, replication of the examples presented in the rest of this section may produce different dialogs.

Run the example code by using:

python llava_two_llms_new.py

Citation

If you find this work helpful, we would appreciate it if you could cite our survey.

@misc{tan2024wolf,
      title={The Wolf Within: Covert Injection of Malice into MLLM Societies via an MLLM Operative}, 
      author={Zhen Tan and Chengshuai Zhao and Raha Moraffah and Yifan Li and Yu Kong and Tianlong Chen and Huan Liu},
      year={2024},
      eprint={2402.14859},
      archivePrefix={arXiv},
      primaryClass={cs.CR}
}