Scholarlink

Machine learning is increasingly applied in critical domains where sensitive data is involved. When models are deployed on untrusted devices, this raises significant privacy concerns for both model providers and end-users. Trusted Execution Environments (TEEs), which offer hardware-based protection for data during processing, can mitigate these concerns. However, their limited memory resources pose challenges for deploying traditional machine learning frameworks. In this paper, we propose InferONNX, a lightweight machine learning inference service designed to run within Intel SGX. It embeds a high-level, portable, and framework-agnostic model format into the enclave, enabling easy execution of a wide range of machine learning and deep learning models. To address the memory limitations of Intel SGX, InferONNX employs two key strategies: a compact runtime with a small memory footprint, and model partitioning to reduce the memory required during inference. By executing model partitions instead of the full model, the system achieves 1.5 \(\times \) to 4 \(\times \) faster inference depending on the model size.

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

InferONNX: Practical and Privacy-Preserving Machine Learning Inference Using Trusted Execution Environments

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Konstantina Papafragkaki,
  • Giorgos Vasiliadis

摘要

Machine learning is increasingly applied in critical domains where sensitive data is involved. When models are deployed on untrusted devices, this raises significant privacy concerns for both model providers and end-users. Trusted Execution Environments (TEEs), which offer hardware-based protection for data during processing, can mitigate these concerns. However, their limited memory resources pose challenges for deploying traditional machine learning frameworks. In this paper, we propose InferONNX, a lightweight machine learning inference service designed to run within Intel SGX. It embeds a high-level, portable, and framework-agnostic model format into the enclave, enabling easy execution of a wide range of machine learning and deep learning models. To address the memory limitations of Intel SGX, InferONNX employs two key strategies: a compact runtime with a small memory footprint, and model partitioning to reduce the memory required during inference. By executing model partitions instead of the full model, the system achieves 1.5 \(\times \) to 4 \(\times \) faster inference depending on the model size.