Scholarlink

This study investigates electromagnetic fault injection (EMFI) on lightweight neural networks hardware. Three compact neural networks (MobileNet, ResNet, EfficientNet) were trained first on Fashion-MNIST and later with CIFAR-10 datasets, then implemented on a custom NVM-based lightweight tinyML hardware platform for EMFI susceptibility testing. We demonstrate that the stored model weights on NVM, can be corrupted by EM injection on the NVM chip during network loading. Further, we demonstrate that the EMFI corrupted weights can lead upto 40% reduction in inference accuracy in case of highly sensitive lightweight models.

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

Investigation of EM Fault Injection on Emerging Lightweight Neural Network Hardware

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bhanprakash Goswami,
  • Reejit Chetry,
  • Chithambara Moorthii J,
  • Manan Suri

摘要

This study investigates electromagnetic fault injection (EMFI) on lightweight neural networks hardware. Three compact neural networks (MobileNet, ResNet, EfficientNet) were trained first on Fashion-MNIST and later with CIFAR-10 datasets, then implemented on a custom NVM-based lightweight tinyML hardware platform for EMFI susceptibility testing. We demonstrate that the stored model weights on NVM, can be corrupted by EM injection on the NVM chip during network loading. Further, we demonstrate that the EMFI corrupted weights can lead upto 40% reduction in inference accuracy in case of highly sensitive lightweight models.