arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2605.04992cs.CR
收藏

你若懈怠,便会错失:通过神经权重转换自动恢复安全对齐

You Snooze, You Lose: Automatic Safety Alignment Restoration through Neural Weight Translation

University of Pavia(帕维亚大学);Radboud University(拉德堡德大学);University of Zagreb(萨格勒布大学)

Marco Arazzi, Vignesh Kumar Kembu, Antonino Nocera, Stjepan Picek, Saraga Sakthidharan

AI总结:

针对公共LoRA适配器安全对齐缺失或微调导致领域知识退化的问题,提出NeWTral方法,在参数空间内转换不安全适配器,可将平均攻击成功率从70%降至13%,保留88%知识保真度,无需原始训练数据和重新训练。

摘要:

公共代码库如今分发着数千个专用的低秩适配(LoRA)模块,但第三方适配器通常不具备负责任训练时应有的拒绝行为。通过在安全数据上微调来恢复该适配器会引发相反的问题:适配器发布时所提供的领域专业知识会退化。下载已完成适配器的从业者既没有领域语料库,也没有安全语料库,因此无法采用上述两种补救措施。我们提出神经权重转换(NeWTral),它完全在参数空间内将不安全的领域适配器映射到其安全对齐的对应适配器,同时保留其专业知识。NeWTral是一种在不安全到安全适配器对上预训练的非线性转换器,具有一个混合专家路由,每个张量都融合了高保真的外科手术式转换器和激进的对齐专家。在四个架构系列(Llama、Mistral、Qwen、Gemma)、最高720亿参数规模以及八个专业领域上,NeWTral将平均攻击成功率从70%降至13%,同时保留了88%的知识保真度。该效果在JailbreakBench、从公共模型中心下载的14个适配器中的12个以及自动越狱攻击下均得到复现。修复仅需对权重进行一次遍历,无需原始训练数据,也无需重新训练。

英文摘要:

Public repositories now distribute thousands of specialized Low-Rank Adaptation (LoRA) modules, but a third-party adapter routinely arrives without the refusal behavior the same adapter would have had if it had been trained responsibly. Restoring it by fine-tuning on safety data induces the opposite failure: the domain expertise the adapter was published to provide degrades. A practitioner who downloads a finished adapter holds neither the domain corpus nor a safety corpus, so neither remedy is available. We propose Neural Weight Translation (NeWTral), which maps unsafe domain adapters to their safety-aligned counterparts entirely within parameter space while preserving their expertise. NeWTral is a non-linear translator pre-trained on unsafe-to-safe adapter pairs, with a Mixture of Experts router that blends a high-fidelity surgical translator and an aggressive alignment expert per tensor. Across four architectural families (Llama, Mistral, Qwen, Gemma) at scales up to 72B parameters and eight professional domains, NeWTral cuts the average Attack Success Rate from 70% to 13% while retaining 88% knowledge fidelity. The effect reproduces on JailbreakBench, on 12 of 14 adapters downloaded from a public model hub, and under automated jailbreak attacks. Curing is a single pass over the weights, needing no original training data and no retraining.

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。

↑