A Resource-Efficient Fault Tolerance Method for Edge-Side DNNs Based on OVSF Coding

JIN Song; WU Bingshuo

doi:null

您当前的位置：

首页 >

文章列表页 >

A Resource-Efficient Fault Tolerance Method for Edge-Side DNNs Based on OVSF Coding

更新时间：2026-05-11

- A Resource-Efficient Fault Tolerance Method for Edge-Side DNNs Based on OVSF Coding
- Telecommunications Science (2026)
- 作者机构：
  
  1.华北电力大学电子与通信工程系，河北保定 071003
  2.华北电力大学河北省电力物联网技术重点试验室，河北保定 071003
  3.华北电力大学电力物联智慧化技术河北省工程研究中心，河北保定 071003
- 作者简介：
- 基金信息：
  
  The Science and Technology Program of Hebei(SZX2020034);Natural Science Foundation of Hebei Province(F2021502006)
- DOI：
  CLC： TP393
- Received：08 January 2026，
  
  Revised：2026-02-15，
  
  Accepted：20 March 2026，
- 稿件说明：
移动端阅览
JIN Song, WU Bingshuo. A Resource-Efficient Fault Tolerance Method for Edge-Side DNNs Based on OVSF Coding[J/OL]. Telecommunications Science, 2026.
DOI：

JIN Song, WU Bingshuo. A Resource-Efficient Fault Tolerance Method for Edge-Side DNNs Based on OVSF Coding[J/OL]. Telecommunications Science, 2026. DOI：

摘要

针对资源受限边缘设备上神经元级容错方法阈值存储开销过大的问题，报告提出了一种基于正交可变扩频因子（Orthogonal Variable Spreading Factor，OVSF）基码的神经元级阈值压缩方法。该方法将激活阈值表示为少量OVSF正交基码的线性组合，仅需存储稀疏系数及其索引，并在推理阶段通过在线重构机制恢复完整阈值，从而在不引入显著计算开销的前提下显著降低存储与传输需求。实验结果表明，在AlexNet与VGG16等典型模型上，该方法在阈值保真度保持在94%以上且容错性能与未压缩FitAct方案基本一致的前提下，将阈值存储开销降低50%–90%。进一步的FPGA(Field Programmable Gate Array)边缘平台实验验证了该方法在带宽受限场景下的可实现性与效率优势。该工作为在边缘设备上部署高效、低开销的容错神经网络提供了一种可行方案。

Abstract

To address the excessive storage overhead of neuron-level fault-tolerance methods on resource-constrained edge devices

this paper proposes a neuron-level threshold compression method based on Orthogonal Variable Spreading Factor (OVSF) basis codes. In the proposed approach

activation thresholds are represented as linear combinations of a small number of orthogonal OVSF basis codes. Only sparse coefficients and their indices are stored

while full thresholds are reconstructed on-the-fly during inference through a lightweight online reconstruction mechanism. This design significantly reduces threshold storage and transmission overhead without introducing substantial computational cost. Experimental results on representative DNN models

including AlexNet and VGG16

show that when the threshold fidelity score (TFS) is maintained above 94%

the proposed method achieves fault-tolerance performance comparable to the uncompressed FitAct scheme. Under this condition

the threshold storage overhead can be reduced by 50%–90%

depending on the selected compression ratio. Furthermore

FPGA-based edge experiments verify the feasibility and efficiency of the proposed method under bandwidth-constrained deployment scenarios. These results indicate that the proposed approach provides a practical and resource-efficient solution for deploying low-overhead fault-tolerant neural networks on edge devices.

关键词

Keywords

references

BOJARSKI M , DEL TESTA D , DWORAKOWSKI D , et al . End to end learning for self-driving cars [J/OL ] . arXiv: 1604.07316 , 2016 .

ESTEVA A , KUPREL B , NOVOA R A , et al . Dermatologist-level classification of skin cancer with deep neural networks [J ] . Nature , 2017 , 542 ( 7639 ): 115 - 118 .

LI G , PATTABIRAMAN K , CHERNIARUBAN S C , et al . Understanding the vulnerability of deep neural networks to soft errors [J ] . ACM Transactions on Embedded Computing Systems (TECS) , 2018 , 17 ( 2 ): 25 .

MUKHERJEE S . Soft errors in modern electronic systems [C ] // IEEE International Symposium on Defect and Fault Tolerance in VLSI Systems . 2008 : 3 - 11 .

BAUMANN R . Soft errors in advanced computer systems [J ] . IEEE Design & Test of Computers , 2005 , 22 ( 3 ): 258 - 266 .

HE Y , ZHU C , SAVARIA Y , et al . Evaluating the robustness of deep neural networks against bit-flip errors [C ] // 2019 IEEE 25th International Symposium on On-Line Testing and Robust System Design (IOLTS) . 2019 : 159 - 164 . DOI: 10.1109/IOLTS.2019.8854389 http://dx.doi.org/10.1109/IOLTS.2019.8854389 .

KOREN I , KRISHNA C M . Fault-tolerant systems [M ] . San Francisco : Morgan Kaufmann , 2020 .

GHAVAMI B , SADATI M , FANG Z , et al . FitAct: error resilient deep neural networks via fine-grained post-trainable activation functions [C ] // 2022 Design , Automation & Test in Europe Conference & Exhibition (DATE) . 2022 : 1239 - 1244 . DOI: 10.23919/DATE54114.2022.9774635 http://dx.doi.org/10.23919/DATE54114.2022.9774635 .

CHEN Z , LI G , PATTABIRAMAN K . A low-cost fault corrector for deep neural networks through range restriction [C ] // Proceedings of 2021 51st Annual IEEE/IFIP International Conference on Dependable Systems and Networks . 2021 : 1 - 13 . DOI: 10.1109/DSN48987.2021.00018 http://dx.doi.org/10.1109/DSN48987.2021.00018 .

CHEN H , LI G , KHAN H M , et al . CLIP-Act: an accurate and efficient fault tolerance scheme for deep neural networks [C ] // 2019 IEEE 37th International Conference on Computer Design (ICCD) . 2019 : 361 - 369 .

ZHANG Z , LI G , PATTABIRAMAN K , et al . A survey on fault tolerance techniques for deep neural networks [J ] . Journal of Computer Science and Technology , 2021 , 36 ( 4 ): 906 - 930 .

郭宇辉 , 闫亚旗 , 付韬等 . 边缘算力发展态势分析 [J ] . 电信科学 , 2025 , 41 ( 11 ): 1 - 13 . DOI： 10.11959/j.issn.1000-0801.2025195 http://dx.doi.org/10.11959/j.issn.1000-0801.2025195 .

KOOPMAN P . Tesla's “Full Self-Driving” hardware (and other production autonomy claims) [EB/OL ] . ( 2017-02-06 )[ 2025-12-12 ] .

MAHMOUD A , AKRAMI M , LI G , et al . Selective node hardening for fault-tolerant deep neural networks [C ] // 2019 IEEE International Symposium on Defect and Fault Tolerance in VLSI and Nanotechnology Systems (DFT) . 2019 : 1 - 6 .

ZHOU C , DU X , YAN M , et al . SAR: Sharpness-aware minimization for enhancing DNNs’ robustness against bit-flip errors [J ] . Journal of Systems Architecture , 2024 , 138 : 103284 .

XU H , LIAO L , LIU X , et al . Fault-tolerant deep learning inference on CPU–GPU integrated edge devices with TEEs [J ] . Future Generation Computer Systems , 2024 , 161 : 404 - 414 .

ZHENG W , XU B , GU J , et al . SAVE: Software-implemented fault tolerance for model inference against GPU memory bit flips [C ] // Proceedings of the 2025 USENIX Annual Technical Conference (USENIX ATC) . 2025 .

XUE X , LIU C , MIN F , et al . ApproxABFT: Approximate algorithm-based fault tolerance for neural network processing [J ] . IEEE Transactions on Very Large Scale Integration (VLSI) Systems , 2025 (Early Access).

XUE X , LIU C , MIN F , et al . Adaptive soft error protection for neural network processing [EB/OL ] . arXiv : 2407 . 19664 v 2 , 2025 .

VITERBI A J . CDMA: Principles of spread spectrum communication [M ] . Reading : Addison-Wesley , 1995 .

LENG J , ZHANG S , LI G , et al . OVSF-inspired compressing of convolutional neural networks [C ] // Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI) . 2018 : 2383 - 2389 .

PAN H , HAMDAN E , ZHU X , et al . Multi-channel orthogonal transform-based perceptron layers for efficient ResNets [EB/OL ] . arXiv : 2303 . 06797 v 2 , 2024 .

HAMDAN E , CETIN A E . HTMA-Net: Towards multiplication-avoiding neural networks via Hadamard transform and in-memory computing [EB/OL ] . arXiv : 2509 . 23103 v 2 , 2025 .

KIM S , SHIN J , WOO S , et al . HOT: Hadamard-based optimized training [C ] // Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 2025 .

HORADAM K J . Hadamard matrices and their applications [M ] . Princeton : Princeton University Press , 2007 .

STRANG G . Introduction to linear algebra [M ] . 5th ed . Wellesley : Wellesley-Cambridge Press , 2016 .

KRIZHEVSKY A , SUTSKEVER I , HINTON G E . ImageNet classification with deep convolutional neural networks [C ] // Advances in Neural Information Processing Systems 25 (NIPS 2012). 2012 : 1097 - 1105 .

Views

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

AI engineering technology solutions in network intelligence

Related Author

Mingwei ZHU

Related Institution

No data

⁰