arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2511.04834cs.LGcs.AIcs.CV
收藏

基于提示的安全引导对已遗忘有害概念的文本到图像扩散模型无效

Prompt-Based Safety Guidance Is Ineffective for Unlearned Text-to-Image Diffusion Models

KAIST(韩国科学技术院)

Jiwoo Shin, Byeonghu Na, Mina Kang, Wonhyeok Choi, Il-Chul Moon

AI总结:

针对文本到图像扩散模型中遗忘有害概念方法与基于提示的安全引导结合时性能下降的问题,提出用概念反演获得的隐式负嵌入替换负提示的方法,在裸体和暴力基准上提升了防御成功率且保留输入语义。

摘要:

文本到图像生成模型的最新进展引发了担忧,即当输入恶意文本提示时,这些模型可能生成有害内容。为解决此问题,出现了两种主要方法:(1)微调模型以遗忘有害概念;(2)利用负提示的无训练引导方法。然而,我们观察到,结合这两种正交方法往往导致防御性能微乎其微甚至下降。这一观察表明两种范式之间存在严重的不兼容性,阻碍了它们的联合有效性。在这项工作中,我们通过提出一种概念简单但实验稳健的方法来解决此问题:将无训练方法中使用的负提示替换为通过概念反演获得的隐式负嵌入。我们的方法不需要修改任何一种方法,可轻松集成到现有 pipeline 中。我们在裸体和暴力基准上实验验证了其有效性,证明在保持输入提示核心语义的同时,防御成功率得到了持续提升。

英文摘要:

Recent advances in text-to-image generative models have raised concerns about their potential to produce harmful content when provided with malicious input text prompts. To address this issue, two main approaches have emerged: (1) fine-tuning the model to unlearn harmful concepts and (2) training-free guidance methods that leverage negative prompts. However, we observe that combining these two orthogonal approaches often leads to marginal or even degraded defense performance. This observation indicates a critical incompatibility between two paradigms, which hinders their combined effectiveness. In this work, we address this issue by proposing a conceptually simple yet experimentally robust method: replacing the negative prompts used in training-free methods with implicit negative embeddings obtained through concept inversion. Our method requires no modification to either approach and can be easily integrated into existing pipelines. We experimentally validate its effectiveness on nudity and violence benchmarks, demonstrating consistent improvements in defense success rate while preserving the core semantics of input prompts.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑