LLM-Assisted Red Teaming of Diffusion Models through "Failures Are Fated, But Can Be Faded"

Som Sagar, Aditya Taparia, Ransalu Senanayake

NeurIPS Workshop on Red Teaming GenAI, 2024

picture_as_pdf PDF
Overview figure for the red teaming work, showing a language model scoring mismatches between prompts and diffusion model outputs to produce a reward signal.

Summary

This extension of Failures Are Fated applies the framework to text-to-image diffusion models, using LLM-assisted generation of rewards and states and reinforcement learning search strategies to make discovery of failure modes more efficient.

BibTeX

@inproceedings{sagar2024llm,
    title={LLM-Assisted Red Teaming of Diffusion Models through "Failures Are Fated, But Can Be Faded"},
    author={Sagar, Som and Taparia, Aditya and Senanayake, Ransalu},
    booktitle={Red Teaming GenAI: What Can We Learn from Adversaries?},
    year={2024}
}