LLM-Assisted Red Teaming of Diffusion Models through “Failures Are Fated, But Can Be Faded”
In one sentence. An extension of the Failures Are Fated framework to text-to-image diffusion models with LLM-generated rewards and states, action screening inspired by design of experiments, and a comparison of DQN, PPO, and A2C for red teaming. SummaryThis work extends the Failures Are Fated framework to red teaming of text-to-image diffusion models. We introduce LLM-generated rewards and states, action screening inspired by design of experiments to reduce the search space of candidate failure factors, and a comparison of DQN, PPO, and A2C as the underlying reinforcement learning algorithms for discovering failure modes in generative models. Note. Builds on Failures Are Fated, But Can Be Faded (ICML 2024 spotlight). BibTeX@inproceedings{sagar2024llmredteaming,
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 = {NeurIPS Workshop on Red Teaming GenAI: What Can We Learn from Adversaries?},
year = {2024}
}
Topicsred teaming, diffusion models, text-to-image, generative model safety, reinforcement learning, LLM-assisted evaluation |