Privacy-Preserving Blockchain-Based Federated Learning for IoT Devices
背景
Home appliance manufacturers strive to obtain feedback from users to improve their products and services to build a smart home system. To help manufacturers develop a smart home system, we design a federated learning (FL) system leveraging a reputation mechanism to assist home appliance manufacturers to train a machine learning model based on customers’ data. Then, manufacturers can predict customers’ requirements and consumption behaviors in the future.
主要贡献
1.分层结构：将深度学习中的全连接层交给MEC（移动边缘计算节点），减少了用户本地的计算。
2. 提出新的归一化方法，同时利用差分隐私阻止了攻击者利用模型进行推理的攻击。
3. 设计了基于区块链的联邦学习系统，通过可追究的方式阻止恶意用户注入垃圾参数。
系统设计
系统分为生产厂商、拥有设备的用户、区块链
生产厂商
对于任务/初始模型的发布，文章给了三种方式：1. 具有随机选择参数的模型存储在区块链中（本文采用）；2. 由厂商发布给所有用户；3.存储在第三方。
由于本文使用区块链去中心化，因此厂商并不会直接参与到训练过程。
用户
Step 1 (Customers Download the Initial Model From the Blockchain)：用户从区块链下载联邦学习初始模型
Step 2 (Customers Extract Features on the Mobile)：用户通过手机收集智能设备数据，收集完成后开始训练数据。由于直接干扰原始数据会损害模型的准确性，因此作者将卷积神经网络（CNN）层视为特征提取器，然后再将噪声注入提取后的数据（疑问：此方案只能适用于CNN神经网络吗？）
Step 3 (Customers Train Fully Connected Layers in the MEC Server)：用户将原始标签和注入噪声的特征发送给MEC server，在MEC端训练全连接层，最后将结果发送回用户
Step 4 (Customers Upload Models to the Blockchain)：迭代步骤2和步骤3，训练结束后，用户使用自己的私钥对模型进行签名，通过手机用户将模型和签名一并发送给区块链
区块链
区块链：consortium blockchain
一致性协议：Algorand（based on Poof of Stake and Byzantine fault tolerance）
1）Miners通过竞争的方式推选leader，资产越多的Miners越有机会成为leader 2）随机选取的委员会验证leader产生的block，只有2/3的成员签名并同意才能通过该block 3）委员会成员执行gossip protocol协议将新块随机广播给相邻节点
文章提出，由于区块的大小受限，因此作者采用了IPFS方法作为off-chain存储手段。以下对实现的一些细节作补充：
首先，miner会验证用户上传的数字签名，只有有效的数字签名能将交易放入交易池（transaction pool） 其次，被随机选取的miners组成委员会负责检测池子里的的交易是否合法，文章采用的方法是Multi-KRUM
Miners Verify the Validity of the Uploaded Model
Algorand采用了一种可以实现无交互的可验证随机函数VRFs，以此选择委员会并根据优先级产生leader，值得一提的是VRFs的机制实现了财产（Stake）越多，被选择为委员会和leader的概率就越高 假设miner A拥有m个币，所有miners总共拥有M个币，现在需要选取 t 个miners作为委员会，那么每个币被选择到的概率 p 就为 t / M。 对于A而言，首先用自己的私钥、一个种子seed，以及标识抽签阶段的一些数据通过VRFs计算出一个hash值和证明，[0, 1]整数区间被划分为m+1个子区间，而被选择的委员会成员就根据 hash / (2^hashlen) 的值决定（hashlen代表hash的长度，式子能将随机的hash值得到一个区间在[0, 1]的结果），而leader就在票数最高的委员会成员中产生
个人理解：j 为 A 所抽到的票（币）数，概率满足二项分布
A Selected Leader Updates the Model
激励机制
为了激励更多的用户参与到联邦学习中，作者结合Multi-KRUM在预防poisoning attack的同时设计了激励机制。
a = H denotes a high evaluation result while a = L denotes a low evaluation result and h means average reputation of whole customers
由于差分隐私个人没怎么研究过，因此文章中的差分隐私创新部分就省略了
附录：
consortium blockchain
联盟链是指在由某个组织内部选定多个预选的节点为记账人，获得大部分所有预选节点的认可才会生成新的区块，其他接入节点可以参与交易，但不过问记账过程（本质上还是托管记账，只是变成分布式记账，预选节点的多少，如何决定每个块的记账者成为该区块链的主要风险点），其他任何人可以通过该区块链开放的API进行限定查询
gossip protocol
参考链接：流言算法/流行病协议
Multi-KRUM
一种拜占庭节点检测机制，每一轮服务器算出每两个用户的欧几里得度量，将得分最低（凑在一块）的一定用户数选取为可信节点
KRUM首次提出是根据梯度进行计算，后面衍生出根据权值(参数)进行欧式距离的计算
It’s perfect time to make some plans for the future and it’s time to be happy.
I have read this post and if I could I desire
to suggest you few interesting things or suggestions.
Maybe you could write next articles referring to this article.
I desire to read even more things about it!
I am willing to listen
I’ve been exploring for a bit for any high-quality articles or weblog posts in this
sort of area . Exploring in Yahoo I finally stumbled upon this
site. Studying this info So i’m glad to exhibit that I’ve a very just right uncanny
feeling I discovered exactly what I needed. I so much definitely will make
certain to don?t disregard this website and give it a look on a constant basis.
Fantastic post however , I was wanting to know if you could write a
litte more on this topic? I’d be very thankful if you could elaborate a little bit more.
Cheers!
Thank you for your affirmation and encouragement. I was in my first year of grad school when I wrote this, but now I have to face the pressure of finding a job, so most of my energy is now spending in preparing for that. Maybe in the future I can take a break, then I will consider continuing to update related articles.
Can I simply say what a comfort to find someone that actually understands what theyre discussing on the net. You actually realize how to bring a problem to light and make it important. More and more people really need to read this and understand this side of your story. I was surprised that youre not more popular since you definitely possess the gift.
I sincerely appreciate your appreciation. What you said that give me a lot of encouragement.
Thankfulness to my father who informed me regarding this blog, this webpage is in fact amazing.
Thanks for your blog, nice to read. Do not stop.
I am impressed with this internet site, very I am a big fan .
Bel article, je l’ai partagé avec mes amis.
Nice post, na-share ko na sa mga kaibigan ko.
Buen post, lo he compartido con mis amigos.
Гарний пост, я поділився ним із друзями.
Dejligt indlæg, jeg har delt det med mine venner.
Bel article, je l’ai partagé avec mes amis.
After looking into a few of the blog articles on your site, I
seriously appreciate your technique of writing a blog.
I saved it to my bookmark webpage list and will be checking back in the near
future. Take a look at my web site too and tell me what you think.
It iis pedrfect tike to ake some plan for thhe longer
tertm andd it’s time too be happy. I’ve learn this submit andd iff I
could I wish to suggest you some fascinating things orr tips.
Maybe you can write subequent articles referring too this article.
I wan to learn moore issues approximately
it!