Gradient-based Automatic Attack of Text-to-Image Model

Abstract

Recently, text-to-image models have been thriving. Despite their powerful generative capacity, our research has uncovered a lack of robustness in this generation process. Specifically, the introduction of small perturbations to the text prompts can result in the blending of primary subjects with other categories or their complete disappearance in the generated images. In this thesis, we propose \textbf{Auto-attack on Text-to-image Models (ATM)}, a gradient-based approach, to effectively and efficiently generate such perturbations. By learning a Gumbel Softmax distribution, we can make the discrete process of word replacement or extension continuous, thus ensuring the differentiability of the perturbation generation. Once the distribution is learned, ATM can sample multiple attack samples simultaneously. These attack samples can prevent the generative model from generating the desired subjects without tampering with the category keywords in the prompt. ATM has achieved a 91.1\% success rate in short-text attacks and an 81.2\% success rate in long-text attacks. Further empirical analysis revealed four attack patterns based on the following: 1) the variability in generation speed, 2) the similarity of coarse-grained characteristics, 3) the polysemy of words, and 4) the positioning of words.