Groot: Adversarial Testing for Generative Text-to-Image Models with Tree-based Semantic Transformation

• DOI: 10.48550/arxiv.2402.12100
• Publication Date: 2024-02-19
• Authors: Yi Liu; Guowei Yang; Gelei Deng; Feiyue Chen; Yuqi Chen; Ling Shi; Tianwei Zhang; Yang Liu
• Keywords: Tree (set theory)

With the prevalence of text-to-image generative models, their safety becomes a critical concern. adversarial testing techniques have been developed to probe whether such models can be prompted produce Not-Safe-For-Work (NSFW) content. However, existing solutions face several challenges, including low success rate and inefficiency. We introduce Groot, first automated framework leveraging tree-based semantic transformation for models. Groot employs decomposition sensitive element drowning strategies in conjunction with LLMs systematically refine prompts. Our comprehensive evaluation confirms efficacy which not only exceeds performance current state-of-the-art approaches but also achieves remarkable (93.66%) on leading as DALL-E 3 Midjourney.