LoS Trick: Countermeasure against CPA for DNN Models Using Loss of Significance in Multiply-Accumulate Operations
Authors:
- 26 November 2024
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 240-246
Abstract
The recent advancements in deep neural network (DNN) models have revolutionized image recognition, natural language processing, and healthcare. However, deploying DNN models in IoT devices introduces significant security risks, particularly side-channel attacks such as correlation power analysis (CPA). In this paper, we introduce a novel countermeasure named loss of significance (LoS) trick LoS trick is designed to protect DNN models from CPA attacks by reducing signal-to-noise ratio (SNR) during the multiply-accumulate operations. We reduce information leakage in power consumption traces by adding large absolute values in the initial stages of multiply-accumulate operations. LoS trick increases the difficulty of obtaining accurate weight parameters. This paper details the theoretical foundation of LoS trick and its implementation methodology and validates its effectiveness through simulations. The results demonstrate that LoS trick outperforms conventional countermeasures, providing a scalable solution to enhance the security of DNN models in IoT environments.Keywords
This publication has 9 references indexed in Scilit:
- Side-Channel Fuzzy Analysis-Based AI Model Extraction Attack With Information-Theoretic Perspective in Intelligent IoTIEEE Transactions on Fuzzy Systems, 2022
- Shuffling Countermeasure against Power Side-Channel Attack for MLP with Software Implementation2021
- Model Reverse-Engineering Attack against Systolic-Array-Based DNN Accelerator Using Correlation Power AnalysisIEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, 2021
- Systolic Array Based Accelerator and Algorithm Mapping for Deep Learning Algorithms2018
- In-Datacenter Performance Analysis of a Tensor Processing Unit2017
- ImageNet classification with deep convolutional neural networksCommunications of the ACM, 2017
- From NLP (Natural Language Processing) to MLP (Machine Language Processing)2010
- Correlation Power Analysis with a Leakage Model2004
- Why systolic architectures?Computer, 1982