Backdooring Bias into Text-to-Image Models

원문 초록

Text-conditional diffusion models, i.e. text-to-image, produce eye-catching images that represent descriptions given by a user. These images often depict benign concepts but could also carry other purposes. Specifically, visual information is easy to comprehend and could be weaponized for propaganda -- a serious challenge given widespread usage and deployment of generative models. In this paper, we show that an adversary can add an arbitrary bias through a backdoor attack that would affect even benign users generating images. While a user could inspect a generated image to comply with the given text description, our attack remains stealthy as it preserves semantic information given in the text prompt. Instead, a compromised model modifies other unspecified features of the image to add desired biases (that increase by 4-8x). Furthermore, we show how the current state-of-the-art generative models make this attack both cheap and feasible for any adversary, with costs ranging between $12-$18. We evaluate our attack over various types of triggers, adversary objectives, and biases and discuss mitigations and future work. Our code is available at https://github.com/jrohsc/Backdororing_Bias.