HE同态加密(Microsoft SEAL) 实例小试

HE 在人脸识别中的应用demo 演示文档

说明

1. 使用的HE方案是CKKS

/
| Encryption parameters :
|   scheme: CKKS
|   poly_modulus_degree: 8192
|   coeff_modulus size: 200 (60 + 40 + 40 + 60) bits
|   scale: pow(2.0,40)
\

2. 输入数据的长度 n = 10 ,
3. 因为没有用FaceNet数据集的缘故,演示所用的测试数据来源为简单生成,生成方式为:
4. 编译平台: Visual Studio 2019
5. 所需外部环境: seal.h ; examples.h ； bits/stdc++.h
6. 项目代码上传至: github: https://github.com/sizaif/SEALExamples/tree/main

数据来源

输入的input数据和database数据来源生成代码如下:

// slot_count = poly_modulus_degree /2  => 4096 
double curr_point = 0;
double step_size = 1.0 / (static_cast<double>(slot_count) - 1); 
ofstream out;
out.open("database.txt", ios::in | ios::out | ios::binary | ios::trunc);
if (out.is_open()) {
   
    int step = 0;
    for (size_t i = 0; i < slot_count; i++)
    {
   
        // 每10行一个数据
        if(step %10 == 9)
            out << curr_point << "\n";
        else {
   
            out << curr_point << " ";
        }
        step++;
        curr_point += step_size;
    }
    out << endl;
    out.close();
    cout << "Input vector: " << endl;
}

选取的input数据如下:

[0.984127 0.984371 0.984615 0.98486 0.985104 0.985348 0.985592 0.985836 0.986081 0.986325]

从生成的database中最后10行中选取第6行生成

---

选取的database E 数据如下:

[0.974359 0.974603 0.974847 0.975092 0.975336 0.97558 0.975824 0.976068 0.976313 0.976557]
[0.976801 0.977045 0.977289 0.977534 0.977778 0.978022 0.978266 0.97851 0.978755 0.978999]
[0.979243 0.979487 0.979731 0.979976 0.98022 0.980464 0.980708 0.980952 0.981197 0.981441]
[0.981685 0.981929 0.982173 0.982418 0.982662 0.982906 0.98315 0.983394 0.983639 0.983883]
[0.984127 0.984371 0.984615 0.98486 0.985104 0.985348 0.985592 0.985836 0.986081 0.986325]
[0.986569 0.986813 0.987057 0.987302 0.987546 0.98779 0.988034 0.988278 0.988523 0.988767]
[0.989011 0.989255 0.989499 0.989744 0.989988 0.990232 0.990476 0.99072 0.990965 0.991209]
[0.991453 0.991697 0.991941 0.992186 0.99243 0.992674 0.992918 0.993162 0.993407 0.993651]
[0.993895 0.994139 0.994383 0.994628 0.994872 0.995116 0.99536 0.995604 0.995849 0.996093]
[0.996337 0.996581 0.996825 0.99707 0.997314 0.997558 0.997802 0.998046 0.998291 0.998535]

---

过程步骤

总览

1. 首先计算 p 与 E 中的 $C_i$ 比较分数并用 $r_i$表示,求最佳的 $r_i$,

   ​ ${\mathbf{r}}_i=\mathrm{dist}\left(\mathbf{p},{\mathbf{c}}_i\right)\cdot \mathbf{k}$ ; k = { 1 , 0 , . . . 0 } k=\left \{ 1,0,...0 \right \} k={ 1,0,...0}

2. 处理完全部数据库内容后,N个单独结果的vector $\mathbf{R}=\left[\begin{array}{c}{\mathbf{r}}_1\\ {\mathbf{r}}_2\\ ⋮\\ {\mathbf{r}}_N\end{array}\right]=\left[\begin{array}{c}\left({\mathbf{r}}_{1,1},0,\dots 0\right)\\ \left({\mathbf{r}}_{2,1},0,\dots 0\right)\\ ⋮\\ \left({\mathbf{r}}_{N,1},0,\dots 0\right)\end{array}\right]$ 每个$r_i$是单独加密的, 可以随机打乱$r_i$的顺序

3. 通过将$r_i$移位,变成对角矩阵 例如 R = [ ( r 1 , 1 , 0 , … 0 ) ( 0 , r 2 , 1 , … 0 ) ⋮ ( 0 , 0 , … r N , 1 )