发表机构
Department of Computer Science Virginia Tech; Verisign, Inc.（弗吉尼亚理工大学计算机科学系; 威瑞信公司）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
针对文本到图像扩散模型中的组合概念擦除，提出无训练的层级接地语义手术框架，通过层级跨度接地和动态属性绑定实现精准擦除，显著降低层级规避和属性泄漏，达到最先进性能。
AI 中文摘要
从已部署的文本到图像扩散模型中移除受版权保护、不安全或用户指定的概念，如今已成为一项实际需求。权重编辑方法可以抑制固定目标，但需要针对每个目标进行重新训练并修改模型检查点。另一方面，无训练方法部署友好，但在组合提示上存在文本侧路由失败的问题。在此类提示中，擦除目标可能通过相关类别而非其词汇名称被调用，且其修饰词可能迁移到保留对象上。本文提出层级接地语义手术（HGSS），一种用于组合概念擦除的无训练框架。该框架提升了文本侧擦除所使用的路由信号和编辑算子。首先，层级跨度接地通过词汇、分类学和语义证据解析擦除目标跨度，同时防范宽泛上位词和复合中心词误报。其次，动态属性绑定通过反事实参考和保留感知的交叉注意力目标，在早期去噪阶段细化文本条件，保持存活的属性-名词绑定完整。HGSS在不更新模型权重或添加学习参数的情况下选择性移除擦除目标。在SEE上，HGSS将层级规避从29.54降至10.02，并将成对属性泄漏大致减半，在报告的擦除方法中取得了最佳的Neighbor E和AttrP分数。在UnlearnCanvas上，HGSS将六指标平均值较匹配的语义手术基线略有提升，达到最先进水平。
英文摘要
Removing copyrighted, unsafe, or user-specified concepts from a deployed text-to-image diffusion model is now a practical requirement. Weight-editing methods can suppress fixed targets, but they require per-target retraining and modify the model checkpoint. Training-free methods, on the other hand, are deployment-friendly, but they suffer from text-side routing failures on compositional prompts. In such prompts, the erase target may be invoked through a related class rather than its lexical name, and its modifiers may migrate onto preserved objects. This paper proposes Hierarchically Grounded Semantic Surgery (HGSS), a training-free framework for compositional concept erasure. The framework lifts both the routing signal and the edit operator used by text-side erasure. First, hierarchical span grounding resolves erase-target spans through lexical, taxonomic, and semantic evidence, while guarding against broad-hypernym and compound-head false positives. Second, dynamic attribute binding refines the text conditioning during early denoising via a counterfactual reference and a preserve-aware cross-attention objective, keeping surviving attribute-noun bindings intact. HGSS selectively removes the erase target without updating model weights or adding learned parameters. On SEE, HGSS cuts hierarchical evasion from 29.54 to 10.02 and roughly halves pairwise attribute leakage, achieving the best Neighbor E and AttrP scores among the reported erasure methods. On UnlearnCanvas, HGSS slightly improves the six-metric average over the matched Semantic Surgery baseline, reaching state-of-the-art.
CommentsAccepted at BMVC 2026
