Pysyft框架

一 简介

Syft是 OpenMined 的open source堆栈，提供Python 中secure的private数据科学。Syftprivate data使用联合学习、差分隐私和加密计算等技术将模型训练与模型训练分离。这是通过类似numpy的界面和与Deep Learning框架的集成来完成的，因此您Data Scientist可以在使用这些新的privacy-enhancing techniques.

Syft允许 aData Scientist询问questionsadataset并在 的privacy limits集合内data owner获取answers这些questions，所有这些都无需获取copy数据本身的 a 。我们称这个过程Remote Data Science。这意味着在domains整个社会中，与某人risks共享信息（copying数据）的潮流，例如隐私侵犯、IP 盗窃和勒索，将不再阻碍benefits安全访问将提供的大量创新、见解和科学发现。

数据拥有方：

1. 部署domain server
2. 上传数据到 自己的domain server中
3. 创建账户，加入网络到 organisation domain server中

数据分析统计:

1. 安装syft,grid,
2. 通过账户连接domain

import syft as sy
domain_client = sy.login(
       port=8081,
       email="[email protected]",
       password="changethis"
    )

3. 寻找当前domain下的数据，和其他domian下的数据进行建模，训练等
   
   mpc协议目前支持
   

二 使用

1. 上传数据在domian中，建立数据使用账户,数据拥有方上传数据

import pandas as pd
import syft as sy
 
#load data
canada_data = pd.read_csv("../datasets/ca - feb 2021.csv")[0:40000]
canada_data.head()
 
#loggin domain
# Let's login into the domain
ca = sy.login(email="[email protected]", password="changethis", port=8081)
 
# update data
# We will upload only the first 40k rows and three columns
# All these three columns are of `int` type
sampled_canada_dataset = sy.Tensor(canada_data[["Trade Flow Code", "Partner Code", "Trade Value (US$)"]].values)
sampled_canada_dataset.public_shape = sampled_canada_dataset.shape
 
ca.load_dataset(
    assets={"feb2020-40k": sampled_canada_dataset},
    name="Canada Trade Data - First 40000 rows",
    description="""A collection of reports from Canada's statistics
                    bureau about how much it thinks it imports and exports from other countries.""",
)
 
# create user
ca.users.create(
    **{
        "name": "Sheldon Cooper",
        "email": "[email protected]",
        "password": "bazinga",
        "budget":10
    }
)
 
# Accept/Deny Requests to the Domain
 
ca.requests.pandas
ca.requests[-1].accept()

2. 数据分析师连接domain,使用连接domian下的数据

import pandas as pd
import syft as sy
import numpy as np
sy.logger.remove()
 
# Logging into the domain Nodes
 
# We will login into Canada and Italy domain node
ca = sy.login(email="[email protected]", password="bazinga", port=8081)
it = sy.login(email="[email protected]", password="bazinga", port=8082)
 
ca_data = ca.datasets[0]['feb2020-40k']
it_data = it.datasets[0]['feb2020-40k']
 
result = ca_data + it_data
 
# 发送数据查看请求
result.request("I'd like to see the result of the sum of imports/exports across italy and canada.")
 
# 查看最终结果
result.get()

3. mpc mul 示例

import syft as sy
sy.logger.remove()
import numpy as np
data = sy.Tensor(np.array([1,2,3],dtype=np.int32))
 
# 登陆
gryffindor = sy.login(email="[email protected]",password="changethis",port="8081")
slytherin =  sy.login(email="[email protected]",password="changethis",port="8082")
hufflepuff = sy.login(email="[email protected]",password="changethis",port="8083")
 
#同步数据
tensor_1 = data.send(gryffindor)
tensor_2 = data.send(slytherin)
tensor_3 = data.send(hufflepuff)
 
 
mpc_1 = tensor_1 + tensor_2
mpc_2  = tensor_2 + tensor_3
mpc3 = mpc_1 + mpc_2 + 3
 
mpc3.block.reconstruct()
 
# output
array([ 7, 11, 15], dtype=int32)

三 部署

前置环境依赖

We will be setting up the following dependencies before PySyft and PyGrid:

Python >=3.9

pip

Conda

Jupyter notebook

Docker

pip install jupyterlab
# 启动 jupyter lab
 
pip install syft hagrid -i https://mirrors.aliyun.com/pypi/simple/
 
# Congrats for making it this far! You only have one final step remaining, before you unleash the power of Hagrid! The final step is to launch a domain # node, which is as easy as:
hagrid launch 
 
# stop domain
hagrid land 

1.quickstart

https://github.com/OpenMined/PySyft/blob/dev/notebooks/quickstart/00-quickstart.ipynb

2. 检查domain,syft,docker,及启动测试环境容器

https://github.com/OpenMined/PySyft/blob/dev/notebooks/quickstart/01-install-wizard.ipynb

3. 部署domain server

https://github.com/OpenMined/PySyft/blob/dev/notebooks/quickstart/data-owner/00-deploy-domain.ipynb

4. 上传数据到domain server

https://github.com/OpenMined/PySyft/blob/dev/notebooks/quickstart/data-owner/01-upload-data.ipynb

5. 测试mpc simple

https://github.com/OpenMined/PySyft/blob/dev/notebooks/smpc/Simple%20SMPC.ipynb

四 其他:

1. 社区活跃，文档视频等比较多，大多都是面向使用者文档，算子，架构等介绍较少
2. 部署方式集成比较高，需要装的都需要grad 工具来装，大多需要版本依赖等，不能手动解决其安装问题
3. 依赖的包比较大，安装时间比较久
4. 在启动domain server时，需要保证docker正常自行build 镜像并启动，端口正常，版本正常，目前在服务器和mac都无法正常启动
5. !hagrid launch test_domain domain to docker:8081 --tag=latest --tail=false --silent,发现pysyft其中存在部署bug,帮忙顺便给修了