arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.19563cs.CV
收藏

SaTPO:基于过渡偏好优化的文本到图像安全对齐

SaTPO: Text-to-Image Safety Alignment via Transition Preference Optimization

Institute of Information Engineering, Chinese Academy of Sciences(中国科学院信息工程研究所);University of Chinese Academy of Sciences(中国科学院大学);University of Science and Technology Beijing(北京科技大学);Nanyang Technological University(南洋理工大学);Beihang University(北京航空航天大学);Sun Yat-sen University(中山大学)

Ziyi Fang, Jingzhi Li, Xiaojun Jia, Xingxing Wei, Wenqi Ren, Haijun Zhang

AI总结:

该研究提出SaTPO方法,通过在去噪过程中逐步对齐文本到图像扩散模型以实现安全对齐,构建CoProPV数据集,在Stable Diffusion v1.5和SDXL上实验提升了安全-质量权衡,降低了不安全检测率并改善了FID。

摘要:

文本到图像扩散模型的安全对齐旨在抑制有害内容,同时不损害有用的生成能力。然而,对完整图像的偏好将多维生成因素压缩为粗糙信号,可能导致超出安全要求的变化。为解决该问题,我们提出安全过渡偏好优化(SaTPO),其在去噪过程中执行逐步对齐。SaTPO利用以下见解:不安全提示及其安全重写在共享父状态和噪声实现的条件下,自然形成紧密耦合的偏好对。通过将优化锚定在相同的父状态上,SaTPO消除了历史轨迹分歧,将安全适应限制在增量对齐中,同时保留与任务相关的语义。多尺度漂移约束和良性重放进一步限制行为偏差,保留普通生成。我们还构建了CoProPV,这是一个经过视觉验证的对比的优化数据集。在Stable Diffusion v1.5和SDXL上的实验表明,安全-质量权衡得到改善。与AlignGuard相比,Inappropriate Image Prompts上的PerspectiveVision不安全检测率分别从12.70%降至6.00%、从15.04%降至9.30%,同时FID从32.28降至25.95、从36.50降至27.51。警告:本文包含潜在冒犯性内容。

英文摘要:

Safety alignment of text-to-image diffusion models aims to suppress harmful content without compromising useful generation capabilities. However, preferences over completed images compress multidimensional generation factors into a coarse signal, potentially driving changes beyond those required for safety. To address this, we propose Safety Transition Preference Optimization (SaTPO), which enforces step-wise alignment during denoising. SaTPO leverages the insight that an unsafe prompt and its safe rewrite naturally form a tightly coupled preference pair when conditioned on a shared parent state and noise realization. By anchoring optimization on identical parent states, SaTPO eliminates historical trajectory divergence, confining safety adaptation to an incremental alignment while preserving task-relevant semantics. Multi-scale drift constraints and benign replay further limit behavioral deviation and preserve ordinary generation. We additionally construct CoProPV, a refined dataset ensuring visually verified contrasts. Experiments on Stable Diffusion v1.5 and SDXL demonstrate improved safety-quality trade-offs. Compared with AlignGuard, PerspectiveVision unsafe-detection rates on Inappropriate Image Prompts decrease from 12.70% to 6.00% and from 15.04% to 9.30%, respectively, while FID improves from 32.28 to 25.95 and from 36.50 to 27.51. Warning: this paper includes potentially offensive content.

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。

↑