ESC

Type to search articles...

No articles found.

↑ ↓ Navigate ↵ Open
Esc Close
Blog Tags GitHub
All tags

confidential-computing

2 posts tagged with "confidential-computing"

July 4, 2026

CloakLM: Obfuscating GPU Memory Layout to Mitigate Model Ex-filtration for Serving

针对第三方共享加速器基础设施上的模型外泄威胁(如 Hermes 通过被动 PCIe 观测无损重建 DNN、TunnelS 通过驱动级访问以高吞吐外泄 HBM),提出软件层内存混淆框架 CloakLM。结合 PCIe 流量整形、层间/层内权重乱序与物理 HBM 页面重映射三种机制,让授权执行保持有效虚拟内存布局且开销可忽略,而未授权观察者看到的是碎片化、语义不连贯的状态。集成于 vLLM 与 PyTorch,与机密计算互补,在 LLaMA/Qwen 分布式推理上实现接近原生的性能同时大幅提高对 PCIe 窃听与 HBM dump 攻击的抵抗力。

inference-serving llm-inference confidential-computing system-optimization
July 4, 2026

The Serialized Bridge:Blackwell GPU 机密计算下的 LLM 服务性能恢复

系统性刻画 Intel TDX + NVIDIA GPU-CC 机密计算下 CVM–GPU 桥被串行化的性能规律,并给出可迁移的运行时/加载器/KV 恢复策略;同时首次公开 B300 机密多 GPU NVSwitch 租户能力

inference-serving llm-inference system-optimization kv-cache confidential-computing

Navigation

  • Work

Resources

  • Lexington Themes.

Socials

  • @Mike_Andreuzza
© 2025 MicroStudio. All rights reserved.

MicroStudio is not affiliated with Stripe, Breeew, Astro, or Tailwind Labs, nor is it endorsed or sponsored by them.