SpinDoctor: 偏った画像生成を引き起こす無害なプロンプトの発見

原文アブストラクト

We introduce SpinDoctor, a method for discovering innocuous prompts for text-to-image models that result in images with a chosen target bias, while retaining the quality of the images and prompts. These prompts present a risk for user manipulation, in that a user who is given these seemingly unbiased prompts might be willing to use them and would receive biased images in return. Our method is flexible: it uses models in a black-box way only and the target image bias can easily be changed, and so our method can be applied with various motivations and in various scenarios. Our method shows some transferability between models-we saw the same bias emerge in 78.6% of cases-such that a victim who uses a prompt generated via our method using one model still likely receives a biased image from another model. In our experiments, a subset of which we corroborated with human evaluation, we were able to find an innocuous prompt that moved generated images in the direction of the desired bias in all but two of 304 experiments, with an average improvement of 34.8% and 33.5% on two open-source text-to-image models.