跳到正文
研踪PaperTrail
← 研踪 PaperTrail

CVF · CVPR 2026:cvf:3b22200a8df8de42d8cb52aa25060e31 · 记录 #1

Immunizing Models Against Harmful Long-Horizon Fine-Tuning via Contractive Optimization Dynamics

来源日期(仅到年份) 2026
来源记录日期 2026
雷达观测 2026-11-01 23:40
CVF · CVPR 2026 ↗PDF ↗

摘要

Fine-tuning has become the default way to adapt powerful foundation models, but this also enables low-cost repurposing for harmful objectives. Existing immunization methods try to optimize local geometry or simulate short attacker horizons, and penalize observed loss drops. However, in practice, downstream tuners run thousands of updates and overcome these short-horizon defenses.In this paper, we propose CLAMP (Contractive Long-horizon Attacker Mitigation via Progress-bounding), an immunization method that traps harmful fine-tuning by shaping the attacker's optimization dynamics rather than only the initial landscape. Our key idea is to make harmful training locally contractive, making each update smaller than the last. This yields a closed-form bound on the attacker's training beyond the attacker's simulated training steps. We also introduce a Hessian-free directional curvature penalty, to create adversarial landscapes along harmful descent directions. Our bi-level objective minimizes the attacker's predicted improvement from train step zero to infinity. Experiments show our method withstands long-horizon fine-tuning across classification, generative, and autoregressive settings, substantially reduces harmful task adaptation, while preserving benign utility and fine-tuneability.

论文发现出处

这些页面提供论文线索,不代表同行评审或质量认证。

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

发表信息与关联版本

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

平台记录号代表本站观测到的元数据变化,不是作者修订号。来源日期仅到月份或年份时,按该月或该年首日存档。

记录 #1 · 2026 · 最新观测版本

Immunizing Models Against Harmful Long-Horizon Fine-Tuning via Contractive Optimization Dynamics

Fine-tuning has become the default way to adapt powerful foundation models, but this also enables low-cost repurposing for harmful objectives. Existing immunization methods try to optimize local geometry or simulate short attacker horizons, and penalize observed loss drops. However, in practice, downstream tuners run thousands of updates and overcome these short-horizon defenses.In this paper, we propose CLAMP (Contractive Long-horizon Attacker Mitigation via Progress-bounding), an immunization method that traps harmful fine-tuning by shaping the attacker's optimization dynamics rather than only the initial landscape. Our key idea is to make harmful training locally contractive, making each update smaller than the last. This yields a closed-form bound on the attacker's training beyond the attacker's simulated training steps. We also introduce a Hessian-free directional curvature penalty, to create adversarial landscapes along harmful descent directions. Our bi-level objective minimizes the attacker's predicted improvement from train step zero to infinity. Experiments show our method withstands long-horizon fine-tuning across classification, generative, and autoregressive settings, substantially reduces harmful task adaptation, while preserving benign utility and fine-tuneability.

查看此版本解读 →查看此版本原文 ↗