[Note] The block quote part is the blogger’s own thinking.
Title: "Federation [1] Learning vs. Blockchain: Who is the strongest king in the field of "trusted media" technology? 》
[Q1] Federal study, what is "federation"?
In the new wave of the Internet, federated learning and blockchain are the two popular technologies that have received the most attention. Federated learning [Note] is a distributed machine learning technology that protects privacy in big data services, and blockchain is a distributed ledger technology that realizes value transfer in a decentralized network . So the question is, who is the strongest king in the field of trusted media technology?
[Note] Another article said that "the role of federated learning is mainly to solve data islands."
Continue to extend: What is a "data island"?
[Article source] What is Federal Learning ( https://zhuanlan.zhihu.com/p/100688371 )
Ace Technology · Origin
Federated learning was born in the Google Input Method Optimization Project in 2016. There are three types of services in the Internet industry: horizontal federated learning, vertical federated learning, and federated migration learning [2] . In April 2020, as a new production factor, data will be listed as five production factors alongside land, labor, capital, and technology. This means that while the privacy of personal data is protected by law as an inviolable individual right, the open sharing of data is the lifeblood of the development of the digital economy and is the key to China's victory in the next round of new technology competition among countries. Therefore, federated learning technology has become a conventional method in the industry to resolve the contradiction between data privacy and data sharing [3] .
[Question 2] What is the basis for the classification of the three service forms of federated learning? What are the similarities and differences?
[Question 3] How is the conflict resolved?
The earlier blockchain technology was born in the Bitcoin project in 2009. According to the source of the ledger, it is divided into three service forms: digital currency, smart contract, and application platform. Before 2019, blockchain technology grew more brutally in the form of digital currency. Until October 2019, the Political Bureau of the CPC Central Committee conducted collective learning on blockchain technology, affirming the integrated application of blockchain technology in the new It plays an important role in the technological innovation and industrial transformation of the industry-promoting the deep integration of blockchain and the real economy, which is expected to solve the problems of SME loan financing, bank risk control, and departmental supervision; use blockchain technology to explore the digital economy Model innovation is expected to provide motivation for creating a convenient, efficient, fair competition, stable and transparent business environment.
Trusted medium · Reason
To gain such popularity and importance, federated learning and blockchain have an important common feature: credibility . As the saying goes, "the heart is separated by the belly", and it is generally difficult to establish trust between strangers. This is because in the social competition with limited resources, it is human nature to obtain more benefits. People are worried about the loss of benefits due to fraud. The role is self-evident.
In traditional markets, transactions are supervised and executed by authoritative institutions. For example, the government formulated the "Consumer Rights Protection Law" and "Product Quality Law", and penalized violations. That is to say, authoritative institutions provide the role of "trusted media" to escort market transactions, especially for common commodities, with detailed and standardized management plans for details such as quality and price.
In the emerging Internet market, smart terminal devices are developing rapidly in accordance with Moore's Law, optical fiber networks and 5G wireless networks are gradually popularized, and product innovations emerge endlessly. In contrast, authoritative organizations need to go through a long period of investigation and research before they can formulate corresponding regulations. This makes many Internet products unable to obtain the "trusted media" role of authoritative organizations in a short period of time, which makes users afraid to be bold. Use new products. For example, only 7 years after the emergence of Internet electronic payment, authoritative agencies issued payment licenses for some Internet companies, and this has resulted in the ubiquitous and very convenient mobile phone payment form. Nowadays, new products such as peer-to-peer transfer (improving the convenience of cross-border transactions) and Internet big data cooperation (improving user service levels) still lack mature laws and regulations to carry out necessary management and regulation, and there is an urgent need for usable "trustworthy" products. medium".
Federated learning and blockchain are precisely the technology-based "trusted media" born under this background. Federal credible study that used data in the process of cooperation is irreversible [4] transformation of data, even without supervision authorities, private data will not be compromised.
The credibility of the blockchain lies in the use of group consensus [5] and digital signature [6] technologies in the accounting process . Even without the supervision of an authoritative institution, the recorded transactions cannot be tampered with and cannot be denied. Therefore, such a technological "trusted medium" will provide new productivity for the sustained and healthy development of the national economy.
[Q4] How to ensure "irreversible"/non-tamperable?
[Q5] How does the blockchain let the "group" know?
[Q6] What is the "digital signature" of the blockchain?
Federated learning/blockchain application scenarios
In application, both are used in Internet scenarios. The difference is that federated learning is used for personalized user services, while blockchain is used for point-to-point transaction accounting and contracts.
JD Digital has successfully implemented many projects in the field of federal learning. For example, during the epidemic, in order to ensure the rapid recovery and normal development of agricultural production and life after the epidemic, Jingdong City, based on digital gateway technology for federal learning, homomorphic encryption, and data privacy protection, worked with Sichuan on the basis of ensuring that data is not released from the database. The Agricultural and Rural Bureau of Guanghan City, Deyang City, Province has deployed the "Southwest Poultry Fry Trading Information Platform" to help small and medium-sized enterprises dynamically and real-time match poultry fry supply and demand information. Within two weeks of the launch, the fry trading of 1,250,450 birds was completed.
Jingdong Mathematics also uses federal learning for face recognition. In cooperative scenarios (such as selfies versus selfies), when the false detection rate is one in 100,000, the pass rate is as high as 99.96%. When the rate is one in ten thousand, the pass rate is as high as 99.99%.
On the blockchain, JD Digital has applied it in many fields such as quality traceability, digital deposit certificates, credit networks, financial technology, and value innovation.
In terms of quality traceability, the AI+ blockchain traceability application is the first in the industry to create a Pu'er tea blockchain anti-counterfeiting traceability platform, which effectively solves the pain points of the Pu'er tea circulation process. As of May 2020, JD Digital's blockchain anti-counterfeiting traceability platform Zhizhen Chain has over 1.3 billion traceability data on the chain, more than 900 cooperative brands, more than 70,000 products entered, and more than 7 million after-sales users Access to inquiries, covering rich business scenarios such as fresh agriculture, mother and baby, wine, beauty, second-hand goods, luxury goods, cross-border goods, medicine, and supermarket convenience stores.
JD Digits pioneered "tea face recognition" technology
In terms of digital deposit certificates, JD Digital relies on its self-developed blockchain digital certificate deposit platform to realize functions such as credible deposit, automated evidence collection, one-click proof, and infringement warning. It has been applied to electronic contracts and electronic invoices. , Electronic certificates, electronic bills, Internet litigation, copyright protection and other scenarios. Among them, the blockchain electronic contract platform has served over a million contract signings.
List of functions of JD blockchain digital certificate platform
On the credit network, in response to the difficulty of obtaining evidence for disputes in the consumer leasing market, the credit leasing platform "Jingxiaozu" uses blockchain technology to obtain leasing agreement, order data, leasing process and other data in leasing business through automated processes and complete the "chain" "Operation to ensure the openness and transparency of leasing services. This technology is also used by JD Digital in digital identity, corporate general account, credit lease, logistics credit investigation, etc., to provide convenience for business operations and personal life.
In terms of financial technology, JD Digital's blockchain has been used in the field of asset securitization as early as 2017. For example, the first blockchain ABS standardization solution launched in June 2019 can help ABS business participating institutions such as asset owners, plan managers, law firms, rating agencies, accounting firms, and custodian banks, optimize business processes and save money. Time cost improves the efficiency of ABS issuance business.
In addition, JD Digital also officially open-sourced JD Chain, the underlying engine of JD blockchain . JD Chain is designed for general business scenarios of enterprises to meet the needs of enterprise building block customization, allowing enterprises to quickly access the blockchain world, and helping enterprises conduct massive and efficient data governance.
JD Chain key technical characteristics
Application basis
Both technologies require computing nodes with the willingness to collaborate and consensus.
The difference is,
Federated learning requires data between nodes to be complementary . For example, one node stores characteristics of consumption habits, and the other node stores characteristics such as personality and hobbies. Between the nodes consensus for the federal algorithm [7] , by agreement between the federal information exchange agreement, the implementation model training and reasoning.
[Q7] What is the principle of the federation algorithm?
The blockchain requires all nodes to synchronously record all transaction information . For example, account A pays 1 token to account B. A's payment information and signature will be sent to all nodes on the network, and each node will generate a consistent record. The most critical technology to reach agreement in the blockchain network is the consensus algorithm [8] . Consensus algorithm is the key to solving the consistency problem, helping nodes to keep data consistent in a distributed and decentralized blockchain network. Commonly used consensus algorithms such as Proof of Work (POW), Byzantine Fault Tolerance (BFT), Proof of Share Authorization (DPoS), etc.
[Q8] What is the principle of the "consensus algorithm" of the blockchain?
Application goal
The goal of both technologies is to enhance mutual trust between nodes in a decentralized network .
The difference is,
Federated learning aims to realize the privacy protection technology of " data available but not visible " [9] , and to improve the quality of user services by merging and using data from all parties, thereby creating new value . For example, homomorphic encryption is a privacy protection technology. The generated ciphertext is completely different from the plaintext, and the distribution and sorting properties have undergone tremendous changes. This makes the original data "invisible" and the ciphertext can be specified according to the specified rules. Perform calculations to achieve gradient descent algorithm and model optimization, and achieve "availability".
[Q9] What is "available"? What is "invisible"? How to achieve "available but not visible"?
Blockchain aims to ensure that transaction records cannot be tampered with, use consensus algorithms and distributed technology to solve the problem of double payment in decentralized networks, and ultimately realize the value expression and value transfer of the digital world. For example, in the Bitcoin system, account A pays 1 bitcoin to account B, and broadcasts the information to all "miner" nodes. In order to obtain system rewards, "miners" strive to pack this information into a new block, and for Get more rewards to be a witness to the historical block, which makes the information in the blockchain immutable.
Strong alliance · the future
Federated learning and blockchain have a common application foundation, and a trusted network of multi-party cooperation is realized through technical consensus, which has good complementarity. From the perspective of application goals, federated learning aims to create value, while blockchain aims to express and transfer value.
Therefore, there are the following two basic combinations .
The first combination is to use the non-tampering feature of blockchain records to trace and punish malicious attacks that may be faced by federal learning partners . For example, while multiple participants are performing federated learning, a blockchain is deployed to record the data fingerprints of federated learning (including modeling samples, inference samples, and interactive information), and the corresponding raw data is stored locally on the participants. When a sample is found to be maliciously attacked, each participant or a third party will form an investigation team to verify the original data based on the fingerprints recorded by the blockchain, and then you can find out which party was attacked and take corresponding measures. .
The second combination is to use the value expression and transfer function of the blockchain to book and distribute the value created by the federated learning service . For example, while multiple participants are performing federated reasoning, the blockchain is deployed to record the user service interface call log fingerprints, the contributions of each participant, and the revenue generated by the service, and the revenue is automatically distributed to All parties involved. This method is different from the existing billing by bill, which can more accurately evaluate the quality of each call, thereby encouraging participants to ensure the accuracy of the call and actively optimize the effect.
[Note] The main method of information transfer is copying; the main method of information transfer is recording.