基于激活签名的表示感知遗忘：从抑制到实体签名擦除
Representation-Aware Unlearning via Activation Signatures: From Suppression to Entity-Signature Erasure
AI总结：
提出ERUF框架，通过挖掘实体特异性激活签名并抑制对应方向，实现表示层面的遗忘，同时保持表面抑制、内部衰减和效用保留。
摘要：
实体级遗忘通常通过模型输出评估：是否停止命名目标、拒绝查询或改变真值比分布。然而，这些输出级测试无法显示主体的内部表示是否被衰减。我们引入实体表示遗忘框架（ERUF），这是一个表示感知框架，挖掘主体特定的激活签名，抑制相应的激活方向，并将行为蒸馏到LoRA参数中。在评估的基线中，ERUF是唯一同时实现表面级抑制、内部衰减和效用保留的方法。在TOFU forget10上，ERUF达到FQ=0.99和MU=0.62，匹配报告的神谕效用，同时接近神谕遗忘质量。在大多数标准基础模型设置中，ERUF保持低泄漏和低内部目标激活，SMR在0.00%至1.10%之间，EL10低于0.06，效用漂移低于3%。在Llama-3.1-8B上，对抗性实体恢复从63.89%降至20.15%，而名称无关恢复减少72.7%至77.4%。联合表面/内部诊断进一步揭示了推理优先模型中仅靠表面指标无法发现的尺度依赖行为。我们将这些结果解释为表示层面衰减的操作性证据，而非不可逆删除的正式保证。
英文摘要：
Entity-level unlearning is usually evaluated by what a model says: whether it stops naming the target, refuses a query, or shifts a Truth Ratio distribution. These output-level tests, however, do not show whether a subject's internal representation has been attenuated. We introduce the Entity Representation Unlearning Framework (ERUF), a representation-aware framework that mines subject-specific activation signatures, suppresses the corresponding activation direction, and distills the behavior into LoRA parameters. Among evaluated baselines, ERUF is the only method that jointly achieves surface-level suppression, internal attenuation, and utility preservation. On TOFU forget10, ERUF achieves FQ = 0.99 and MU = 0.62, matching reported oracle utility while approaching oracle forget quality. Across most standard foundation-model settings, ERUF maintains low leakage and low internal target activation, with SMR between 0.00% and 1.10%, EL10 below 0.06, and utility drift below 3%. On Llama-3.1-8B, adversarial entity recovery falls from 63.89% to 20.15%, while name-agnostic recovery decreases by 72.7% to 77.4%. Joint surface/internal diagnostics further reveal scale-dependent behavior in reasoning-prior models that surface metrics alone would miss. We interpret these results as operational evidence of representation-level attenuation, not as a formal guarantee of irreversible deletion.
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。
