FHE, or Fully Homomorphic Encryption, allows direct computation on encrypted data, and the computed results are consistent with the results of performing the same computation on the original data. This means that we can process and analyze data without exposing the original data, providing a new solution for protecting data privacy and integrity.

The theoretical foundation of FHE can be traced back to 1978 when Rivest et al. proposed the first homomorphic encryption problem. It wasn't until 2009 that Gentry, in his doctoral thesis, constructed the first realizable fully homomorphic encryption scheme, marking the beginning of the flourishing development of FHE research. Early FHE schemes had extremely low computational efficiency, unable to meet practical application requirements. With researchers continuously proposing optimization schemes, the performance of FHE has been significantly improved, propelling FHE research towards enhancing efficiency, reducing computational complexity, and expanding application scenarios.

Technical Implementation Paths of FHE

Ideal Lattice-based FHE Schemes

This is currently recognized as the most practical and efficient construction method for FHE. The main idea is to leverage the algebraic structure of ring operations, completing encryption and decryption through modular arithmetic and factorization of ideal factors in the ring. Representative schemes include BGV, BFV, and CKKS, among others. The advantage of these schemes lies in their high computational efficiency, while the drawback is the need for larger key and ciphertext spaces.

Matrix-based FHE Schemes

These schemes encode plaintext information into matrices and achieve homomorphic properties through matrix operations. Representative schemes include GSW and HiNC, characterized by high security but relatively low efficiency.

NTRU-based FHE Schemes

NTRU (Number Theory Research Unit) has a good algebraic structure and cyclic symmetry, enabling the construction of efficient FHE schemes such as YASHE and NTRU-FHE. The advantage of these schemes is their relatively small key and ciphertext sizes, making them suitable for resource-constrained environments.

FHE Schemes based on Learning With Errors (LWE/LWR)

These schemes are constructed based on the LWE/LWR hardness problems, such as FHEW and TFHE. These schemes prioritize theoretical innovation and can achieve extremely strong security, but their practicality is limited.

FHE vs. ZKP

FHE and ZKP are both cryptographic techniques, but they are more or less complementary.

ZKP allows a prover to demonstrate to a verifier that a statement is correct without revealing any details about the statement. The verifier can verify the correctness and completeness of the computation without needing to re-execute it. Although ZKP can prove correctness without revealing information, its inputs are typically in plaintext form, which may lead to privacy leaks.

The introduction of FHE solves this problem. FHE enables arbitrary computation on encrypted data without decryption, thereby protecting data privacy. However, FHE cannot independently ensure the correctness and reliability of the computation, which is precisely the problem that ZKP solves.

By combining FHE and ZKP, the advantages of both techniques can be leveraged. FHE protects the privacy of input data and the computation process, while ZKP provides an cryptographic proof of correctness, legality, and auditability for FHE computations. Ultimately, this combination achieves truly secure and trustworthy private computation, which is highly valuable for privacy-preserving applications such as sensitive data processing and multi-party computation.

ZAMA: The FHE Pioneer

Zama is an open-source cryptography company building cutting-edge FHE solutions for blockchain and AI. They offer four major open-source solutions:

TFHE-rs is a Rust implementation of fully homomorphic encryption over the Torus, used for performing Boolean and integer operations on encrypted data. The TFHE-rs library implements Zama's variant of TFHE, which enables all necessary homomorphic operations, such as addition and function evaluation through programmable bootstrapping.

Concrete is an open-source FHE framework that includes a TFHE compiler as part of the framework. This compiler translates regular programming code into executable instructions that a computer can use with FHE, allowing developers to easily write FHE programs. Computations on FHE-encrypted data may introduce substantial noise, leading to errors. Concrete has a very low default error probability, and developers can flexibly modify the parameters of this error probability.

Concrete ML is an open-source Privacy-Preserving Machine Learning (PPML) tool built on Concrete, enabling developers to integrate FHE into machine learning models without requiring cryptographic knowledge.

fhEVM introduces FHE into the EVM ecosystem, allowing developers to execute encrypted smart contracts on-chain while still preserving the composability of smart contracts under the premise of protecting on-chain data privacy. fhEVM integrates TFHE-rs and introduces a new TFHE Solidity library, enabling developers to perform computations on encrypted data using Solidity.

Fhenix: The First FHE-Rollup

Fhenix is the first Layer 2 Rollup based on FHE, built upon Zama's TFHE-rs. It has developed its own cryptographic computation library, fheOS, which includes precompiled opcodes for common encryption operations, enabling smart contracts to use FHE primitives on-chain. fheOS is also responsible for communication and authentication between the rollup and the Threshold Service Network (TSN) for decryption and re-encryption requests, while proving the legitimacy proof of the decryption requests. The fheOS library aims to be integrated as an extension into any existing EVM network, while being fully compatible with the EVM.

Fhenix's consensus mechanism uses Arbitrum's Nitro prover. The reason for choosing a fraud proof approach is that the underlying structures of FHE and zkSNARK are different, and using ZKP to verify FHE computations would be extremely computationally intensive, making it almost impossible to achieve at the current technological stage.

Fhenix has recently collaborated with EigenLayer to develop FHE coprocessors, introducing FHE computations into other public chains, L2s, and L3s. Since Fhenix employs fraud proofs with a 7-day challenge period, EigenLayer's services can help coprocessors achieve fast transaction confirmations, significantly improving performance.

Inco Network: Confidentiality as a Service

Inco is a modular confidential computing Layer 1 that can serve as a universal confidentiality  layer for Web3. It supports fhEVM, enabling developers to rapidly build confidential DApps using the Solidity language and development tools within the Ethereum ecosystem. Additionally, Inco provides Confidentiality as a Service (CaaS) to EVM and Cosmos chains lacking native encryption through bridging and the IBC protocol. The CaaS service primarily offers three functionalities:

On-chain encrypted state: Directly store encrypted data on-chain without off-chain storage.

Composable encrypted state: Perform state transitions on encrypted data entirely on-chain without decryption.

On-chain randomness: Generate randomness for applications on-chain without the need for external randomness services, allowing for direct on-chain application development.

Currently, Inco has several use cases, including gaming, NFTs, RWA, voting governance, and DID.

Mind Network: FHE Restaking Layer

Mind Network is the first FHE Restaking Layer tailored specifically for AI and PoS networks. As a Restaking Layer, it accepts restaking tokens from ETH, BTC, and AI bluechips for staking. Simultaneously, as an FHE validation network, it utilizes FHE technology to validate node data and reach consensus, ensuring data integrity and security. Mind Network provides economic security guarantees for decentralized AI, DePIN, EigenLayer AVS, Babylon AVS, and critical PoS networks, maintaining the overall system's consensus and trustworthiness.

Restaking Layer: Collaborating with EigenLayer, StakeStone, Renzo, Babylon, Ankr, and others, Mind Network derives security from the Ethereum and Bitcoin networks. It also collaborates with Chainlink CCIP, Connext, and others to enable cross-chain remote restaking.

Security Layer: Mind Network introduces FHE-enhanced validators to ensure end-to-end encryption during the validation and consensus computation processes, and further enhance security by integrating the modular fhEVM of Fhenix and Inco.

Consensus Layer: Mind Network introduces a Proof of Intelligence (PoI) consensus mechanism designed specifically for AI tasks, ensuring fair and secure reward distribution among FHE validators. Additionally, Mind Network is collaborating with AltLayer, EigenDA, and Arbitrum Orbit to launch a Rollup chain, enhancing consensus computation with lower costs and faster performance.

Privasea: Proof of Human

Privasea is a DePIN+AI network for FHE machine learning, whose technical architecture includes the following core components:

HESea Library: An advanced FHE library that provides secure computation capabilities on encrypted data. This library supports various FHE schemes such as TFHE, CKKS, BGV/BFV, and others.

Privasea API: The application programming interface for the Privasea AI network, offering a suite of functionalities and endpoints to simplify data submission, model training, and prediction operations, while ensuring data encryption during transmission and processing.

Privanetix: A decentralized computation network composed of multiple high-performance computing nodes capable of efficiently processing encrypted data. Each node integrates the HESea library to ensure data privacy and computational performance.

Privasea Smart Contract Suite: A blockchain-based incentive mechanism that tracks the registration and contributions of Privanetix nodes through smart contracts, validates computations, and distributes rewards, ensuring participant engagement and fairness.

Privasea has launched the ImHuman application, which develops a Proof of Human based on FHE, aiming to prove that users are human beings to protect their digital identities from impersonation by bots and artificial intelligence. Users can verify their human identity through facial biometric recognition technology, generating a unique NFT as proof of their individual human identity. With the ImHuman application, users can securely attest their personal identities on Web3 and Web2 platforms without revealing personal details.

Conclusion

In the DeFi domain, FHE enables transactions and fund flows to take place without disclosing sensitive financial information, thereby protecting user privacy and reducing market risks. It may also serve as one of the effective solutions to the MEV problem. For fully on-chain games, FHE ensures that player scores, game progress, and other data are protected in an encrypted state, while simultaneously allowing game logic to be executed on-chain without exposing the data, enhancing fairness and security of the games. In the AI field, FHE allows for the analysis and model training on encrypted data, not only protecting data privacy but also promoting cross-institutional data sharing and collaboration, driving the development of more secure and compliant AI applications.

While FHE still faces numerous challenges in practicality and efficiency, its unique theoretical foundation brings hope for overcoming these bottlenecks. In the future, FHE is expected to significantly enhance performance and expand application use cases through algorithm optimization, hardware acceleration, and other approaches, providing a more robust foundation for data privacy protection and secure computation.

References

https://github.com/zama-ai/tfhe-rs

https://github.com/zama-ai/concrete

https://github.com/zama-ai/concrete-ml

https://github.com/zama-ai/fhevm

https://www.fhenix.io/introducing-fhe-rollups-scaling-confidential-smart-contracts-on-ethereum-and-beyond/

https://www.fhenix.io/fhe-coprocessors-fhenix-eigenlayer-join-forces-for-next-gen-onchain-confidentiality/

https://docs.inco.org/

https://docs.mindnetwork.xyz/minddocs

https://mindnetwork.medium.com/fhe-secured-restaking-layer-scaling-security-for-ai-depin-networks-73d5c6e5dda3

https://drive.google.com/file/d/1jbxWMgEziupt119gvM1n0Mu8sDdM7VWF/view

https://www.privasea.ai/

Data-Driven: The Solid Foundation of AI and Web3

Data is the core driving force behind AI development, akin to fuel for an engine. AI models require ingesting vast amounts of high-quality data to gain profound understanding and robust reasoning abilities. Data not only provides the training foundation for machine learning models but also determines their accuracy and reliability.

In the traditional centralized AI data acquisition and utilization model, several key issues arise:

• Data acquisition costs are prohibitively high, making it challenging for small and medium-sized enterprises to participate.

• Data resources are monopolized by tech giants, creating data silos.

• Personal data privacy faces risks of leakage and misuse.

Web3 offers a new decentralized data paradigm to address the pain points of traditional models:

• Through projects like Grass, users can sell their idle network capacity to AI companies, enabling decentralized web data crawling, cleaning, and transformation to provide real, high-quality data for AI model training.

• Public AI adopts a "label to earn" model, incentivizing global workers with tokens to participate in data annotation, aggregating global expertise, and enhancing data analysis capabilities.

• Blockchain data transaction platforms like Ocean Protocol and Streamr provide an open and transparent trading environment for data supply and demand, fostering data innovation and sharing.

Nevertheless, real-world data acquisition also faces challenges such as varying data quality, high processing complexity, and insufficient diversity and representativeness. Synthetic data may be the rising star in the Web3 data realm. Based on generative AI technologies and simulations, synthetic data can mimic the properties of real data, effectively complementing it and improving data usage efficiency. In domains like autonomous driving, financial market trading, and game development, synthetic data has already demonstrated its mature application potential.

Privacy Protection: The Role of FHE in Web3

In the data-driven era, privacy protection has become a global focal point, as evidenced by the enactment of regulations like the EU's General Data Protection Regulation (GDPR), reflecting the strict guardianship of personal privacy. However, this has also brought challenges: some sensitive data cannot be fully utilized due to privacy risks, undoubtedly limiting the potential and reasoning capabilities of AI models.

FHE, or Fully Homomorphic Encryption, allows direct computation on encrypted data without the need for decryption, while the computation results are consistent with performing the same operations on plaintext data.

FHE provides robust protection for AI privacy computation, enabling GPU computing power to execute model training and inference tasks without accessing the original data. This presents significant advantages for AI companies, as they can securely open API services while protecting trade secrets.

FHEML supports encrypted processing of data and models throughout the entire machine learning lifecycle, ensuring the security of sensitive information and preventing data leakages. In this way, FHEML reinforces data privacy and provides a secure computing framework for AI applications.

FHEML complements ZKML, where ZKML proves the correct execution of machine learning, while FHEML emphasizes computing on encrypted data to maintain data privacy.

The Computing Revolution: AI Computation in Decentralized Networks

The computational complexity of current AI systems doubles every three months, leading to a surge in computing power demands that far exceeds the supply of existing computing resources. For instance, the training of OpenAI's GPT-3 model required immense computing power, equivalent to 355 years of training time on a single device. Such a shortage of computing power not only limits the progress of AI technology but also renders advanced AI models inaccessible to most researchers and developers.

Additionally, the global GPU utilization rate is below 40%, coupled with the slowdown in microprocessor performance improvements and supply chain and geopolitical factors contributing to chip shortages, exacerbating the computing power supply issue. AI practitioners find themselves in a dilemma: either purchasing hardware or renting cloud resources, desperately needing an on-demand, cost-effective computing service model.

IO.net is a decentralized AI computing power network based on Solana, aggregating idle GPU resources globally to provide AI companies with an economical and accessible computing power market. Entities demanding computing power can publish computation tasks on the network, and smart contracts will allocate tasks to contributing miner nodes. Miners execute the tasks, submit the results, and receive reward points upon successful verification. IO.net's approach improves resource utilization efficiency, helping to alleviate computing power bottlenecks in fields like AI.

In addition to general decentralized computing power networks, there are platforms dedicated to AI training, such as Gensyn and Flock.io, as well as specialized computing power networks focused on AI inference, like Ritual and Fetch.ai.

Decentralized computing power networks provide a fair and transparent computing power market, breaking monopolies, lowering application barriers, and improving utilization efficiency. Within the Web3 ecosystem, decentralized computing power networks will play a crucial role, attracting more innovative dApps and jointly driving the development and application of AI technologies.

DePIN: Web3 Empowering Edge AI

Imagine your smartphone, smartwatch, or even smart home devices possessing the ability to run AI – this is the allure of Edge AI. It enables computation to occur at the data source, realizing low latency and real-time processing while protecting user privacy. Edge AI technology has already been applied in critical domains such as autonomous driving.

In the Web3 realm, we have a more familiar name – DePIN. Web3 emphasizes decentralization and user data sovereignty, and DePIN enhances user privacy protection by processing data locally, reducing the risk of data leakages. Web3's native token economy can incentivize DePIN nodes to provide computing resources, building a sustainable ecosystem.

Currently, DePIN is rapidly developing in the Solana ecosystem, becoming one of the preferred public chain platforms for project deployment. Solana's high throughput, low transaction fees, and technological innovations have provided strong support for DePIN projects. At present, the market capitalization of DePIN projects on Solana exceeds $10 billion, with notable projects like Render Network and Helium Network achieving significant progress.

IMO: A New Paradigm for AI Model Publishing

The concept of IMO (Initial Model Offering) was first introduced by the Ora protocol, tokenizing AI models.

In the traditional model, due to the lack of a revenue-sharing mechanism, once an AI model is developed and released to the market, developers often struggle to obtain continuous revenue from the subsequent use of the model. Especially when the model is integrated into other products and services, it becomes challenging for the original creators to track usage and, consequently, generate revenue. Additionally, the performance and effectiveness of AI models often lack transparency, making it difficult for potential investors and users to evaluate their true value, limiting market acceptance and commercial potential.

IMO provides a novel method of funding and value-sharing for open-source AI models. Investors can purchase IMO tokens to share in the subsequent revenue generated by the model. Ora Protocol utilizes the ERC-7641 and ERC-7007 ERC standards, combined with an Onchain AI Oracle and OPML technology, to ensure the authenticity of AI models and enable token holders to share in the revenue.

The IMO model enhances transparency and trust, encourages open-source collaboration, aligns with crypto market trends, and injects momentum into the sustainable development of AI technology. While IMO is still in its early experimental stage, as market acceptance and participation expand, its innovative nature and potential value are worth anticipating.

AI Agents: A New Era of Interactive Experiences

AI Agents can perceive their environment, engage in independent thinking, and take appropriate actions to achieve predefined goals. With the support of large language models, AI Agents not only understand natural language but can also plan, make decisions, and execute complex tasks. They can function as virtual assistants, learning user preferences through interactions and providing personalized solutions. Even without explicit instructions, AI Agents can autonomously solve problems, enhance efficiency, and create new value.

Myshell is an open AI-native application platform offering a comprehensive and user-friendly toolset for configuring bot functionalities, appearances, voices, and connecting to external knowledge bases. It strives to create a fair and open AI content ecosystem, empowering individuals to become super creators by leveraging generative AI technologies. Myshell has trained specialized large language models to make role-playing more humanized. Its voice cloning technology can accelerate personalized AI product interactions, reducing voice synthesis costs by 99%, with voice cloning achievable in just 1 minute. Customized AI Agents created with Myshell can currently be applied to various domains, including video chatting, language learning, and image generation.

In the convergence of Web3 and AI, the current focus is primarily on exploring the infrastructure layer, addressing critical issues such as acquiring high-quality data, protecting data privacy, hosting models on-chain, improving the efficient utilization of decentralized computing power, and verifying large language models. As these infrastructural components gradually mature, we have reason to believe that the fusion of Web3 and AI will give birth to a series of innovative business models and services.

首先需要选择一个支持Optimism的钱包，并添加相应的网络设置；然后，通过安全的跨链桥，你可以将1000 USDT和一些用作gas费的ETH桥接至Optimism，进而在相应的DEX连接钱包并执行交易。整个过程虽然直观，但涉及多次交互和等待确认时间，以及潜在的网络费用和交易滑点等不确定因素。对于链上交易新手来说，每一步都可能是挑战，这也限制了链上应用的大规模推广。那么，能否将链上交互的复杂性降至与中心化交易所（CEX）操作同等的简单级别呢？

Intent-centric即“以意图为中心”，不同于前述场景，intent-centric只关注结果，不关注过程。用户只需要发出“我想要在Optimism购买$OP“的指令并签名，Intent-centric协议会协同第三方完成所有必要步骤，从而降低操作门槛，为用户带来类似Cefi的无缝体验。

目前已有多个项目围绕intent-centric进行探索。

Bob the Solver

Bob the Solver 是基于意图交易的基础设施，可以集成到账户抽象钱包和去中心化应用中。在ETHGlobal Paris黑客松表现亮眼，并获得了1inch的最佳用例奖。它由两部分构成：

Solver(求解器): Solver配备ML来识别用户意图，并对意图进行分类，规划实现意图的最优路径。确定路径后，solver构建实现用户意图所需交易并转发至账户抽象钱包。

**账户抽象钱包(AA钱包)：**负责交易的执行，由bundler和paymaster组成。其中，bundler负责接收并规划solver发送的交易，而paymaster则负责管理与这些交易相关的gas费。

Anoma

Anoma是一个以意图为中心的隐私保护协议，用于去中心化交易对手发现、解析和多链原子计算(atomic settlement)。核心流程包括以下几个步骤：

1. 用户交互：用户可以向Anoma的intent gossip网络（P2P网络层）发送透明、私有或有保护的意图。

2. 意图收集与匹配：网络中的solver（可以是searcher等角色）收集意图，进行平衡状态转换，以实现交易对手的匹配。

3. 交易处理：匹配后的交易提交到加密的mempool，validator将proposer打包的区块发送到执行层，完成执行并验证有效性，最终完成状态根更新。

Anoma的执行层构建在Taiga上，这是一个基于Halo2的zk-circuit，使得Anoma可以用以部署以太坊的zk-rollup，共享以太坊的部分安全性。

除了zk-rollup，Anoma还可以用于op-rollup、NFT Marketplace、Dex等的完全去中心化开发。它提供了一种简化的方式来构建和操作这些应用，同时保护用户的隐私和安全。

Essential

Essential专注于构建基于意图的基础设施和工具，以解决MEV（矿工可提取价值）问题。其核心目标是减少MEV所带来的负外部性，使用户能够以最低的成本实现意图。Essential计划建立一个solver网络，捕获用户的意图，并竞争寻找最佳解决方案，将以获取利润为目的的searcher转化为寻求用户最佳解决方案的solver。

为了实现这一目标，Essential正在开发一系列产品：

意图表达的通用领域特定语言(DSL)：实际上是标准化的意图语言，使solver更容易理解和推理用户意图。

以太坊和EVM的intent-centric的账户抽象标准：Solver需要执行链上操作的权限，因此需要账户抽象。新标准将引入ERC-4337，委托solver构建有效交易以满足意图。

模块化意图层：为解决intent-centric架构与底层兼容时不可避免的设计牺牲。其好处包括：

• 简化架构：Intent- only放弃transaction的概念，使solver专注于提供解决方案，无需理解意图与事务之间的依赖；

• 订单流聚合：所有订单通过同一个solver网络分发，确保透明并利用多个流动性来源优化用户结果；

• MEV阻力：激励solver更多地将价值回馈给用户，以提供最佳结果；

• 模块化设计：允许在多个堆栈和生态系统中部署协议，促进跨链意图执行。

Flashbots Suave

值得一提的是，Flashbots推出的suave正是聚焦于基于intent-centric的跨链MEV提取。正如之前提到的，suave的架构围绕用户交易偏好组成，分别为对偏好的表达、执行及结算。这里的“偏好”可以理解为带有条件和约束的用户意图。意图的执行者充分竞争，返还部分MEV收益给用户，以使用户利益最大化。

上个月，flashbots推出了MEVM，为MEV提供了新的预编译，允许任何集中式MEV基础设施转换为去中心化区块链上的智能合约。此外，MEVM通过将敏感数据的计算移动到链下运行的执行节点来提供隐私和效率。MEVM将大大丰富suave链生态，带来更多订单流，用户将获得更好的结果，提议者将获得更优质的区块。

Intent-centric未来有哪些机会？

**意图语言与AI的结合：**大多数情况下，用户的意图是复杂的，意图表达是不充分的，这可能导致solver对意图理解更困难，从而无法规划出最优解决路径。可利用AI更精确识别用户意图，例如利用机器学习构建意图识别模型，根据用户的交易请求来源、交易数据等推断潜在的目的和需求。

**第三方执行层：**Solver识别意图后，会把交易过程打包给第三方执行层。执行过程往往会涉及复杂的交易流程及多方参与者，能够快速响应、准确处理交易流程且安全、稳定的第三方执行层，是实现用户意图的关键之一。但也要警惕第三方执行层的中心化问题及DOS攻击等。

**Defi的大规模采用：**其实当前已有Defi项目采用intent-centric，例如Cowswap、1inch Fusion、UniswapX等。Cowswap推出Cow Hooks以进行基于意图的AMM交换；UniswapX允许用户签署意图，solver在链下执行订单并在链上结算。未来随着intent-centric在Defi领域的大规模采用，有可能改善资本效率和流动性提供。

**全链游戏的发展：**全链游戏的全部游戏逻辑及数据等都以智能合约的形式存储在链上，受限于区块链速度、可扩展性等瓶颈，目前无法支持游戏策略复杂、画面精美等的游戏。全链游戏通过intent-centric可以降低玩家操作的复杂度，玩家只需发出意图指令并签名，即可完成游戏操作，而不是对意图操作的每一步都approve并签名，提高玩家游戏体验。

随着Web3不断发展，用户体验和交互效率逐渐成为焦点，intent- centric使我们能够以全新的方式构建和操作复杂系统，期待一个更加用户友好、效率和透明的去中心化世界。

参考链接：

https://medium.com/anomanetwork/an-overview-of-anoma-s-architecture-26b72e8c9be5

https://blog.essential.builders/introducing-essential/

https://writings.flashbots.net/mevm-suave-centauri-and-beyond

Flashbots推出一系列关于MEV的解决方案，致力于建立一个公平、透明、安全的交易环境，本文将盘点Flashbots为MEV去中心化所做的努力。

Flashbots Auction——矿工去中心化

在PoW时代，MEV的主要参与者包括搜索者和矿工，其中矿工拥有出矿权，并且在交易排序、插队和篡改方面具有特权，从而占据利润的大头。同时，搜索者愿意支付高昂的Gas费用，以确保自己的交易被包含在区块中，某些情况下，搜索者可能会拿出MEV的90%甚至更多作为奖励支付给矿工。

显而易见，这种特权地位对交易的隐私以及安全造成了威胁，带来了包括网络拥堵、Gas竞争等负外部性。

Flashbots Auction 提供了以太坊用户和矿工之间的私密通信渠道，通过透明和公平的方式来协商交易的执行顺序和价格。

• 提交拍卖：搜索者将交易打包成Bundles提交到Flashbots拍卖系统中，同时指定一个最低价格展示交易的潜在价值；

• 矿工竞价：矿工在拍卖系统中对感兴趣的 Bundle 进行竞价，指定自己愿意接受的最低价格；

• 构建区块：矿工从拍卖系统中选择一个或多个竞价最高的Bundle，将其包含在区块中；

• 结算和执行：区块中包含的交易和 Bundle将按照拍卖中达成的顺序和价格进行执行。

通过Flashbots Auction，搜索者不再需要通过Gas War的方式来确保自己的交易被优先打包，也无需支付失败交易的费用，这样使MEV分配更加公平合理，大大提高以太坊网络的安全性和执行效率。

MEV-Boost——验证者去中心化

PoS时代，PoW矿工被验证者替代（在信标链质押32个ETH并运行以太坊客户端即可成为验证者），在每一个Epoch里会随机选择一个验证者作为proposer打包交易并提交至以太坊主网。

Flashbots Auction推出MEV-Boost，引入一个新角色——Builder，负责构建区块。通过PBS（Proposer-Builder Separation）将提议者和构建者分离，促进验证者竞争、去中心化和抗审查。同时验证者接入MEV-Boost，降低寻找利润最大交易的门槛，可大幅增加质押收益。

搜索者通过mempool等渠道寻找有套利机会的交易，将其与自己的交易打包成bundle，发送给区块构建者；

构建者试图挑选最有利可图的bundles打包成区块，发送给Relay；

Relay实际上是为验证者托管构建者打包的区块，relay转发区块头至MEV-Boost，验证者签署区块头来锁定打包区块的承诺后，relay将完整区块转发给验证者；

验证者签署承诺成为区块提议者，负责提议区块到网络中，并将区块添加到链上。

Suave——构建者去中心化

虽然MEV-Boost带来种种好处，但也面临着构建者中心化的问题，当前少数构建者垄断了大部分的区块构建。

构建者中心化的原因主要有两个：

• EOF（独家订单流），例如构建者跟钱包、Dapp等建立合作，通过修改用户默认RPC获取独家订单流；

• 跨域MEV，跨域构建者捕获多链交易，导致整个区块链网络更加中心化。

为此，Flashbots推出Suave，一个高度专业化的即插即用的独立网络，把mempool和构建者与现有的链分开，多链共享同一排序层来保证去中心化。

Suave的架构围绕用户交易偏好组成，分别为对偏好的表达、执行及结算。

通用的偏好表达：Suave的mempool中，公开透明的展示所有EVM链用户的交易。用户在发起交易时，可以表达自己的交易偏好；

最佳的执行市场：执行者监听mempool中的交易，充分竞争提出最优执行价格，执行者会返还部分MEV给用户；

去中心化区块构建：超越单体区块构建者，形成一个去中心化的区块构建网络，网络之间共享订单流和Bundles，在不泄漏交易具体内容的情况下，协作完成区块构建；由于构建者能够打包多链产生的MEV，这也使构建者和验证者能够获得更高的MEV收入。

Relay去中心化

目前在MEV-Boost的生态系统中，Relay中心化的问题还没有完全的解决方案，而Relay在整个MEV链条中扮演着受信任的角色。Flashbots作为市场上最大的Relay运营商之一，对验证者提供免费使用，这导致新的Relay运营商缺乏进入市场的动力。目前Relay高度集中可能会造成例如单点故障、交易审查等问题，部分Relay会与Builder勾结，优先转发合作Builder的区块，拒绝或延迟处理某些正常交易，从而影响市场的充分竞争。

如果Falshbots等头部Relay运营商持续提供补贴，Relay很难形成市场规模以推动去中心化。

值得注意的是，目前Flashbots已不是最大的Relay提供商。Ultra Sound是一个乐观的Relay，无需许可、中立且抗审查。我们很高兴看到Relay多样性的提高，当然这也少不了Flashbots为Ultra Sound、Agnostic Gnosis这样的新Relay提供开源支持。

IOBC已投项目Bloxroute的Relay也接入Flashbots，转发交易至MEV-Boost。Bloxroute BDN网络有分布在全球的Relay，因此使用Bloxroute Relay成功率及速度会更高。2022年，Bloxroute与Flashbots 白帽团队达成合作，允许BNB Chain上的 Flashbots 客户使用BloXroute BDN，实现快速可靠的交易体验。

MEV将“卷”向哪里？

Mempool：有利可图的订单是MEV的起点，这部分创业者主要是卷BD，寻找更多钱包、Dapps等接入自己的RPC，以保护交易隐私并获取独家订单流。有些项目还会通过协议层面将部分MEV收入返还给用户，例如Flashbots的MEV-share。

Builder：这部分创业者主要卷硬件及策略，安全稳定地提供服务，是一个builder被选择的基本因素，而策略则直接影响MEV利润；其中一部分创业者也盯上了跨域MEV的蛋糕，毕竟单链能提取的MEV价值有限。

Validator：验证者的门槛不断降低，成为散户参与MEV的最佳方式之一，针对Validator的创业者主要卷质押ETH的流动性，也就是LSDfi赛道，目前LSDfi赛道也比较拥挤。

未来MEV的发展，仍然面临着一些挑战：

1、跨域MEV。随着Layer2的发展，越来越多的交易从Layer1流向Layer2，但是Rollup的出块及排序是由Sequencer处理，MEV排他性地被Sequencer提取，成为Rollup的重要利润来源之一。共享排序器的出现可能会解决这个问题。

2、MEV利润再分配。目前MEV大部分被验证者提取，各方利益不公平会导致做恶意愿更强。

虽然MEV赛道头部效应明显，但仍然不断吸引更多开发者进入黑暗森林。事实上，MEV还有很多值得深入探讨的地方，例如模块化区块链中MEV的捕获、再质押对MEV供应链的影响等等，也期待能够出现更好的解决方案。