FLOCK 백서

FLOCK

원문

bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 16쪽 중 1쪽

# Introduction to FLock.io FLock.io seeks to decentralise training and [value alignment](https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback). We ensure that AI objectives match the public’s ethics and societal aims, that decision-making falls to communities, and that usefulness is a top priority. FLock knocks down barriers impeding participation in the ecosystem. We allow developers to provide models, data, or compute in a modular way. The result: a plethora of fit-for-purpose models created by, for, and under the stewardship of the communities. Our incentivised platform democratises AI agent training, fine-tuning and inferencing. It halts user data collection, and enables equitable distribution of rewards and widespread governance. Leveraging blockchain technology and AI, FLock provides a powerful environment for dealing with large datasets.&#x20; # The Centralisation Problem This year, Decentralised AI (dAI) has captured the industry spotlight. The [resignation](https://stability.ai/news/stabilityai-announcement) of Stability AI CEO Emad Mostaque highlighted this concern. His departure reflects a growing consensus that AI development should fall under the stewardship of a diverse global community. This followed a series of controversies in the governance of generative AI: deepfakes, the OpenAI boardroom fiasco, Gemini AI's data bias, and the Getty Images lawsuit against Stability AI. #### The problem with centralised control over AI creation The centralisation problem presents an overbearing barrier to AI innovation. Under the status quo, the world’s largest corporations hold sway over the trajectory of AI development based on their own objectives, which do not necessarily align with the public interest. The danger of AI being controlled by centralised corporations is that their biases and values are amplified on a global scale. They decide who gains access to the models, and their value alignment often downgrades the performance of models. We consequently see low public participation, less access to computing power, amplified data bias and inaccuracies from less and lower quality training data, and a missed opportunity for AI to realise its maximal potential as a force for good. There is a pressing need for an equitable distribution of rewards for those who help to refine these models.<br> # Architectural Breakdown {% hint style="success" %} Get started with decentralized training on AI Arena platform: [This way](https://train.flock.io/)! {% endhint %} This section provides a high-level overview of FLock's system, including its system design as well as its blockchain and AI layers. # System Design This section gives a high-level overview of FLock's system design. FLock's system logic is comprised of three major components, namely, AI Arena, FL Alliance and AI Marketplace. <figure><img src="https://742781353-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F1RpcbvTSHzzPwOSUvgKU%2Fuploads%2FYydKabo3VBqQ9GnXF8dv%2Fsystem-design-fig1.png?alt=media&amp;token=9482df79-2f03-4ac3-8099-e84760768960" alt=""><figcaption><p>Figure1. FLock's system logic. </p></figcaption></figure> As shown in Figure 1, upon a task creation, the model is first trained and validated in ***AI Arena***, a blockchain-based decentralised training platform, and then optionally further fine-tuned in ***FL Alliance*** using participants’ local data. Finally, the model is deployed by applications in the ***AI Marketplace***, where feedback will be used to further improve the model. <figure><img src="https://742781353-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F1RpcbvTSHzzPwOSUvgKU%2Fuploads%2FLxZHCOJfAmDJTbxgUeBd%2Fsystem-design-fig2.png?alt=media&amp;token=de36c09a-b7c7-4640-a1ac-e28f1b810b25" alt=""><figcaption><p>Figure2. FLock's system design.</p></figcaption></figure> Specifically, as shown in Figure 2, when a task is first created in AI Arena, it is first trained by training nodes. These nodes then submit their models to validators, who evaluate and propose scores for each submission. The validators reach a consensus on these scores to determine the ranking of the submitted models. The consensus model can then be assigned to FL clients, who fine-tune and improve it using their local data, resulting in the FL global model. The AI Arena consensus model or the FL global model can be deployed and hosted in the AI Marketplace, providing interfaces to various applications. AI Arena training nodes, validators and FL clients need to stake to participate the system, and will be rewarded based on their performance. # Blockchain Layer This section describes how the blockchain layer of FLock enhances both incentivisation and security. ## Incentivisation The blockchain layer plays a pivotal role in reward distribution. It ensures that participants can securely lock in their stakes, fostering an environment of trust and transparency. The process is designed to incentivise participation by allocating rewards based on contributions, thus encouraging a more engaged and active community. Through the use of smart contracts, the system automates the rewards process, making it both efficient and fair. This automation not only reduces the potential for human error but also ensures that rewards are distributed in a timely and fair manner. ## Security At FLock, we propose a generic framework that can integrate an FL system with a blockchain system and can defend against poisoning attacks without adopting complex cryptographic protocols, which is adopted from our peer-reviewed work [here](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=\&arnumber=10471193).&#x20; Specifically, FLock employs a proof-of-stake mechanism with the goal of ensuring robust security and consensus within the network. Such system enhances Byzantine fault tolerance by aligning participants' incentives with network integrity— those who act dishonestly or fail to reach consensus risk losing their staked tokens. This economic disincentive promotes honest behaviour, ensuring the network remains secure and reliable. The table below describes how FLock mitiagtes vairous types of risks which could occur in decentralised training platforms.&#x20; | Attacks | Description | FLock Mitigation | | -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Sybil Attacks | An attacker might gain disproportionate influence in the FLock system by creating and controlling multiple fake identities of participants. | <ul><li>Staking assets increases the difficulty of controlling many training nodes or validators.</li><li>Blind validation mechanism prevents collusion between training nodes and validators.</li><li>In each task, only the top <span class="math">k\_1</span> training nodes and the top <span class="math">k\_2</span> validators will rewarded, ensuring that participants with poor performance do not receive rewards.</li></ul> | | DoS Attacks | An attacker might exhaust the FLock system resource and make it unavailable to honest participants. | <ul><li>Rate limiting is implemented to restrict the frequency and volume of actions within a certain time frame, ensuring that no single participant can overwhelm the system.</li></ul> | | Free-rider Attacks | Free riders benefit from a system without contributing fairly. In the FLock system, a free rider training node may randomly submit models without actuall training. Similarly, free rider validators give random scores instead of honestly evaluating models. | <ul><li>In each task, only the top <span class="math">k\_1</span> training nodes and the top <span class="math">k\_2</span>validators will rewarded, ensuring that participants with poor performance do not receive compensation.</li><li>FLock AI Arena consensus guarantees that honest participants who contribute diligently are appropriately recognised and rewarded, deterring free riders from exploiting the process.</li></ul> | | Lookup Attacks | Training nodes could cheat by learning to predict past validation score calculations. | <ul><li>Two datasets, i.e., Datasets A and B, are used as validation sets to evaluate the models. Consequently, even if a training node manages to optimise its performance for Dataset A, it could still underperform on Dataset B. By carefully calibrating the rewards between Dataset A and B, FLock effectively motivates training nodes towards developing genuinely high-quality models.</li></ul> | | FL Model Poisoning Attacks | In FL Alliance, an attacker may use biased or corrupted data during the training process to degrade the model's performance. | <ul><li>By aggregating contributions, majority voting minimises the impact of single malicious participants.</li><li>The slashing mechanism penalizes malicious clients, deterring model poisoning by reducing their rewards and discouraging future attacks.</li></ul> | # AI Layer ## AI Arena AI layer supports a conventional machine learning (ML) model training paradigm, optimising mod- els directly on users’ devices with their own or public data. To maximise the generalisation ability and performance of the final trained models, this layer is designed to encourage community members to contribute various public or local data, harnessing the broader community’s power. By leveraging blockchain, it ensures contributors are continually engaged and rewarded based on the quantifiable impact of their data on improving the models. ## FL Alliance Ultilising Federated Learning (FL) approach, the AI layer enables thousands of participants to collaboratively train a global model, where data sovereignty is preserved by ensuring that no local data are transmitted at any stage of the training process. Within the AI layer, a model aggre- gation component allows participants to upload weights from models trained on their unique local data. These weights are then aggregated to build an optimal global model, enhancing its generalisation capabilities and performance. The integration of training task automation and deployment orchestration components simplifies the process for users to join tasks and contribute valuable knowledge extracted from their data. ## AI Marketplace Once models are trained and fine-tuned through AI Arena and FL Alliance, they can be hosted on our platform. This platform serves as a comprehensive environment for deploy- ing ML models, making them accessible within blockchain networks of virtual machines (VMs). By integrating with these networks, the platform facilitates the seamless execution and inference of complex ML models, providing real-time, scalable, and secure solutions. # AI Arena Onboarding guide for AI Arena, a pioneer decentralised AI training platform AI Arena is FLock.io's decentralised AI training platform. Through train.flock.io, the AI Arena gateway, users can stake FLOCK, FLock.io's token deployed on Base, in order to participate and earn staking rewards. There are four ways to participate: * Task Creators\*: Define desired models and submit tasks to the platform * Training Nodes: Develop or finetune models required by task creators * Validators: Evaluate the models submitted by training nodes * Delegators: Delegate tokens to training nodes and validators, increasing reward distribution *\*Currently all tasks are created by the FLock.io team. We will open up task creation to the community in the coming months.* # Participants This section describes the roles played by various participants in AI Arena. ## Training Nodes Training nodes are responsible for training models and are required stake tokens to be eligible. This requirement ensures a commitment to the network’s integrity and facilitates a distributed, trust-based mechanism for task assignment. This stake acts both as a gatekeeper to maintain a high standard and as a foundational element in the network’s security protocol, ensuring that nodes have a vested interest in proper execution and the overall health of the ecosystem. ## Validators Validators are responsible for evaluating work done by training nodes, submitting validation scores that influence reward distribution. They participate by staking tokens, which grants them the opportunity to validate tasks assigned to them, ensuring hardware compatibility and fair task distribution proportional to their stake. Upon completion of a task, they can withdraw heir stake and claim rewards, which are calculated based on their performance and adherence to the expected outcomes. The design ensures that validators are incentivised to provide accurate and honest validations, thereby maintaining the quality and reliability of the network’s computational tasks. ## Delegators\* Delegators contribute to the FLock system by supporting other participants’ staking process, enhancing the network’s validation capacity without directly participating in the task training or validation process. Delegators share in the rewards earned by their associated delegatees, based on predefined algorithms that account for their staked contribution. This role allows individuals to participate in the network’s training, validation and economic activities, leveraging their tokens to support delegatees, without needing the technical capabilities to train or validate tasks themselves. Note that training nodes and validators who choose to accept delegation are free to choose a reward share ratio. The higher the ratio, the bigger the reward share their delegators will receive. FLock.io acts solely as a facilitator in the delegation process. The platform provides the technical infrastructure to enable interactions between delegators and delegatees within the ecosystem. FLock.io DOES NOT: * Act as a financial advisor or provide recommendations on delegation decisions. * Influence or control the actions, performance, or decisions of delegators or delegatees. * Guarantee any specific outcomes or returns for delegators. FLock.io disclaims all liability for any losses, damages, or issues arising from: * Mismanagement by delegatees, including, but not limited to, poor decision-making or unethical behavior. * Token slashing penalties incurred by delegatees. * Technical errors, delays, or malfunctions in the delegation process. Delegator Responsibilities: * Risk Acknowledgment: By participating, delegators accept all risks associated with delegation, including the loss of rewards or tokens due to delegatee actions or network vulnerabilities. * Independent Decision-Making: delegators are responsible for selecting delegatees and must conduct their own research before staking tokens. * Compliance: Delegators must comply with all applicable laws and regulations in their jurisdiction. Delegatee Responsibilities: * Reward Share Ratio Transparency: Delegatees must disclose their reward share ratio before accepting stakes and adhere to it consistently. This ratio is entirely determined by the delegatee, and FLock.io does not control or operate it. * Risk Disclosure: Delegatees must inform delegators of potential risks, including slashing penalties or other network-related risks. * Compliance: Delegatees must comply with all applicable laws and regulations in their jurisdiction. \*Note that the delegation function may not be available in some regions due to local regulations. &#x20; # Quickstart Follow the steps in the following sections to participate in AI Arena.&#x20; # Pre-requisites Requirements for decentralised AI training participation ## Set up requirements checklist * [ ] &#x20;A Web3 Wallet - You will use this to connect to the web app at [train.flock.io](https://train.flock.io) * [ ] ETH on Base mainnet for gas fees * [ ] gmFLOCK tokens which can be converted from \[LINK] {% hint style="info" %} After you've joined the platform, head over to our [discord server](https://discord.com/invite/ay8MnJCg2W). To access the developer channels, you can verify your wallet in the [#ai-arena-role](https://discord.com/channels/977200106531082240/1254731167543070831) channel. {% endhint %} # WSL installation To install Conda in WSL (Windows Subsystem for Linux), you can follow these steps: #### Step 1: Install WSL and Set Up a Linux Distribution You can find [Ubunto(WSL) ](https://apps.microsoft.com/detail/9pdxgncfsczv?rtc=1\&hl=en-gb\&gl=GB)in Microsoft store #### Step 2: Download and Install Miniconda or Anaconda You can choose to install either Miniconda (a minimal installer for Conda) or Anaconda (a larger distribution including many scientific packages). **Install Anaconda** 1. Open your WSL terminal. 2. Download the Anaconda installer script using `wget` or `curl`: ```bash wget https://repo.anaconda.com/archive/Anaconda3-latest-Linux-x86_64.sh ``` 3. Run the installer script: ```bash bash Anaconda3-latest-Linux-x86_64.sh ``` 4. Follow the prompts during the installation. Accept the license agreement, specify the installation path (default is usually fine), and choose whether to initialize Anaconda. 5. Initialize Conda (if not done during installation): ```bash source ~/anaconda3/bin/activate ``` 6. Update Conda to the latest version: ```bash conda update conda ``` #### Step 3: Verify Installation 1. Close and reopen your WSL terminal or run the following command to activate Conda: ```bash source ~/.bashrc ``` 2. Verify that Conda is installed and working correctly: ```bash conda --version ``` You should see the Conda version printed in the terminal, indicating that Conda has been successfully installed. By following these steps, you will have Conda installed and ready to use in your WSL environment. # Delegator Guide How to delegate your gmFLOCK to training and validation nodes Delegators contribute to the FLock system by supporting other participants’ staking process, enhancing the network’s validation capacity without directly participating in the task training or validation process. Delegators stake `gmFLOCK`on behalf of other participants, and earn rewards in return.&#x20; {% hint style="info" %} Note that the delegation function may not be available in some regions due to local regulations. {% endhint %} ### Step 0: Go to <https://train.flock.io/stake-to-earn> and connect your wallet. Navigate to <https://train.flock.io/stake-to-earn>. There, connect your wallet. <figure><img src="https://742781353-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F1RpcbvTSHzzPwOSUvgKU%2Fuploads%2FUkeLpN702MfuIKdg1XzC%2Fimage.png?alt=media&amp;token=0556cf25-5fe9-43ad-ac33-46ce251fe07d" alt=""><figcaption><p>Step 0.1: Connect wallet</p></figcaption></figure> You will see a variety of wallets to choose from. Pick your favourite wallet there, agree terms of use and proceed.&#x20; <figure><img src="https://742781353-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F1RpcbvTSHzzPwOSUvgKU%2Fuploads%2FXZ3I3wnwKYuVRQY1h4iY%2Fimage.png?alt=media&amp;token=93499c24-6312-42b3-93fb-b2f10a7953dd" alt=""><figcaption><p>Step 0.2: Choose wallet and agree to terms of use</p></figcaption></figure> {% hint style="info" %} If this is the first time you connect this wallet to train.flock.io, you may need to sign a transaction in order to connect your wallet successfully.&#x20; {% endhint %} ### Step 1: Choose a wallet to delegate to. On the "Stake t
프로젝트로 돌아가기