26、离散和基于事件数据上SNN的安全威胁分析

离散和基于事件数据上SNN的安全威胁分析

1. 生成对抗样本的方法

1.1 算法描述

为了研究SNN和DNN的安全性，提出了生成对抗样本的方法，具体算法如下：

Algorithm 23: Methodology for Generating Adversarial Examples for SNNs and DNNs
1 Given: network (SNN or DNN), original sample X, maximum human perceptual distance Dmax, noise magnitude δ, area A of N · N pixels, number of pixels to perturb M
2 while D(X∗, X) < DMAX do
3     Compute Standard Deviation SD for every pixel of A;
4     Compute Gap(X∗), Gap−(X∗), Gap+(X∗);
5     if Gap(X∗)− > Gap(X∗)+ then
6         VariationPriority(xi,j) =
7             [Gap−(X∗) − Gap(X∗)] · SD(xi,j);
8     else
9         VariationPriority(xi,j) =
10            [Gap+(X∗) − Gap(X∗)] · SD(xi,j);
11    Sort VariationPriority in descending order