Scholarlink

The computing continuum presents a system that enables applications to execute their operations between different edge devices and cloud infrastructure without any interruptions. The continuum shows sensitive computations moving toward regulated privacy-critical data, which needs absolute confidentiality and integrity protection during execution on shared or untrusted platforms. The development of Trusted Execution Environments (TEEs) serves as a fundamental technology that enables this vision through its ability to protect data through hardware-based separation and its remote verification system. The development of portable confidential workloads becomes challenging because TEEs at different stages of the continuum present different security frameworks, verification systems and programming access points. To manage this heterogeneity, the systems community has proposed WebAssembly as a unifying, language-agnostic execution substrate, owing to its strong isolation properties, portability and low runtime overhead. While prior work has explored WebAssembly in CPU-based confidential environments, the recent introduction of VM-based confidential computing support for NVIDIA GPUs raises new questions about how WebAssembly can be integrated with GPU-backed TEEs and how trust can be composed across CPU and accelerator domains. In this paper, we investigate the use of WebAssembly as an abstraction layer for confidential computation across heterogeneous TEEs in the computing continuum. We design and implement a WebAssembly-based architecture that integrates CPU TEE attestation with NVIDIA GPU TEE attestation, enabling a continuous chain of trust across virtual machines and accelerators. We evaluate two WebAssembly-based approaches for executing confidential machine learning workloads on GPU-backed TEEs and compare them against a native baseline. Our experimental results on cloud-based confidential VMs with NVIDIA H100 GPUs demonstrate that WebAssembly can effectively manage TEE heterogeneity without introducing additional attestation overhead, while achieving competitive and in some cases improved performance for secure, GPU-accelerated computation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Approaches for Running WebAssembly Across the Computing Continuum with GPU-Backed TEEs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Giovanni Mazzeo,
  • Fernando Pannullo,
  • Luigi Romano

摘要

The computing continuum presents a system that enables applications to execute their operations between different edge devices and cloud infrastructure without any interruptions. The continuum shows sensitive computations moving toward regulated privacy-critical data, which needs absolute confidentiality and integrity protection during execution on shared or untrusted platforms. The development of Trusted Execution Environments (TEEs) serves as a fundamental technology that enables this vision through its ability to protect data through hardware-based separation and its remote verification system. The development of portable confidential workloads becomes challenging because TEEs at different stages of the continuum present different security frameworks, verification systems and programming access points. To manage this heterogeneity, the systems community has proposed WebAssembly as a unifying, language-agnostic execution substrate, owing to its strong isolation properties, portability and low runtime overhead. While prior work has explored WebAssembly in CPU-based confidential environments, the recent introduction of VM-based confidential computing support for NVIDIA GPUs raises new questions about how WebAssembly can be integrated with GPU-backed TEEs and how trust can be composed across CPU and accelerator domains. In this paper, we investigate the use of WebAssembly as an abstraction layer for confidential computation across heterogeneous TEEs in the computing continuum. We design and implement a WebAssembly-based architecture that integrates CPU TEE attestation with NVIDIA GPU TEE attestation, enabling a continuous chain of trust across virtual machines and accelerators. We evaluate two WebAssembly-based approaches for executing confidential machine learning workloads on GPU-backed TEEs and compare them against a native baseline. Our experimental results on cloud-based confidential VMs with NVIDIA H100 GPUs demonstrate that WebAssembly can effectively manage TEE heterogeneity without introducing additional attestation overhead, while achieving competitive and in some cases improved performance for secure, GPU-accelerated computation.