Primary sourcearXiv: Google DeepMind
View sourceEvent ·Official first-hand
双盲评测:用安全飞地同时保护评测提示词与模型权重
AI安全评测长期存在双向保密难题:评测方不愿公开私有基准提示词,模型方不愿暴露专有模型权重。一篇arXiv论文提出"双盲评测"方案,通过NVIDIA H100安全飞地在硬件层加密RAM数据,并用OpenMined的PySyft维护数据与模型隐私,为双方提供密码学级别的保密保证。 论文团队进行了首次试点:在Google Cloud Platform的H100安全飞地中,用AILuminate基准的私有题库评测Google DeepMind的专有模型Gemini 2.5 Flash Lite。作者称这是首次有技术方案同时实现一方前沿专有模型权重与另一方封闭基准的互相保密。 对第三方评测机构和监管审计而言,这类方案可能降低接触闭源模型时提示词泄露的风险,也为模型厂商接受独立安全审计提供新的技术路径。
Current read
评测机构的私有题库和厂商的模型权重以前总有一方要妥协,这篇论文用H100安全飞地加PySyft让两边都能保密。试点直接跑了Gemini 2.5 Flash Lite和AILuminate私有题库,做安全评测的人应该看看这套流程怎么搭。
11 sources47° AI heatLast updated 2026/11/9 15:48:07
Evidence
11 sourcesRelated source 1marktechpost
View sourceRelated source 2Tech in Asia
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4kimmonismus
View sourceRelated source 5Gary Marcus
View sourceRelated source 6TestingCatalog
View sourceRelated source 7Blognone
View sourceRelated source 8Thomas Wolf
View sourceRelated source 9The Information
View sourceRelated source 10Epoch AI
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.