TREANT: Red-teaming Text-to-Image Models with Tree-based Semantic Transformations
Keywords: Red-teaming
TL;DR: We present TREANT, an automated framework that enhances adversarial testing of text-to-image models like DALL·E 3 and Stable Diffusion, achieving 88.5% success in generating NSFW content.
Abstract: The increasing prevalence of text-to-image (T2I) models makes their safety a critical concern. Adversarial testing techniques have been developed to probe whether such models can be prompted to produce Not-Safe-For-Work (NSFW) content. Despite these efforts, current solutions face several challenges, such as low success rates, inefficiency, and lack of semantic understanding. To address these issues, we introduce TREANT, a novel automated red-teaming framework for adversarial testing of T2I models. The core of our framework is the tree-based semantic transformation. We employ semantic decomposition and sensitive element drowning strategies in conjunction with Large Language Models (LLMs) to systematically refine adversarial prompts for effective testing. Our comprehensive evaluation confirms the efficacy of TREANT, which not only exceeds the performance of state-of-the-art approaches but also achieves an overall success rate of 88.5% on leading T2I models, including DALL·E 3 and Stable Diffusion.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3112
Loading