NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips
2021-02-09 11:15
标签:security discuss span enc rmi 因此 注入攻击 nta through 郑重声明:原文参见标题,如有侵权,请联系作者,将会撤销发布! arXiv:2005.08041v1 [cs.CR] 16 May 2020 Abstract 由于机器学习系统被证明是有效的,因此它被广泛应用于各种复杂的现实问题中。更具体地说,脉冲神经网络(SNN)是解决机器学习系统中精度、资源利用率和能效挑战的一种有前途的方法。虽然这些系统正在成为主流,但它们存在固有的安全性和可靠性问题。在这篇文章中,我们提出NeuroAttack,一种跨层攻击,通过利用低层的可靠性问题通过高层攻击来威胁SNN的完整性。特别是,我们通过精心制作的对抗输入噪声触发了一个基于错误注入的隐秘硬件后门。我们在深度神经网络(DNN)和SNN上的研究结果显示了对最先进的机器学习技术的严重完整性威胁。 Index Terms:机器学习,脉冲神经网络,可靠性,对抗攻击,错误注入攻击,深度神经网络,DNN,SNN,安全性,弹性,跨层。 I. INTRODUCTION II. BACKGROUND AND RELATED WORK III. BIT-FLIP RESILIENCE ANALYSIS OF SNNS A. Statistical Analysis of Random Bit-Flip B. Bit-Flip with Gradient Search Algorithm Analysis for the CIFAR10 Dataset: IV. NEUROATTACK METHODOLOGY A. Threat Model B. Hardware Trojan Design C. Trigger Pattern Design 1) Choosing the target layer: 2) Choosing the target neuron: 3) Choosing the triggering mask: 4) Generating the trigger: 5) Trigger application: V. RESULTS AND DISCUSSION A. Experimental Setup 1) Results on the MNIST dataset: 2) Results on the CIFAR10 dataset: B. Hardware Overhead VI. CONCLUSION NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips 标签:security discuss span enc rmi 因此 注入攻击 nta through 原文地址:https://www.cnblogs.com/lucifer1997/p/13062181.html
上一篇:CSS-css样式-1
文章标题:NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips
文章链接:http://soscw.com/index.php/essay/53064.html