AIArena : A Blockchain-Based Decentralized AI Training Platform

Zhipeng Wang, Rui Sun, Elizabeth Lui, Tuo Zhou, Yizhe Wen, Jiahao Sun · 2024

Read AIArena : A Blockchain-Based Decentralized AI Training Platform by Zhipeng Wang, Rui Sun, Elizabeth Lui, Tuo Zhou, Yizhe Wen, Jiahao Sun (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-14566.txt.

Introduction

A few big firms control most advanced AI training, which concentrates power over data and models. Zhipeng Wang and five co-authors propose AIArena in a 2024 paper, a training platform built on blockchain technology. Anyone can contribute models or computing power, and on-chain rules reward each contribution fairly, aiming for open and joint development.

The team built a working version on a public test network linked to the Base blockchain. Their tests show the platform runs in practice. The authors argue that shared ownership and transparent rewards could widen who gets to shape AI and reduce the bias that comes from closed control.

Catalog record

CREATORZhipeng Wang, Rui Sun, Elizabeth Lui, Tuo Zhou, Yizhe Wen, Jiahao Sun
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-14566

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AIArena : A Blockchain-Based Decentralized AI Training Platform Zhipeng Wang, Rui Sun, Elizabeth Lui, Tuo Zhou, Yizhe Wen, Jiahao Sun Abstract . The rapid advancement of AI has underscored critical challenges in its development and implementation, largely due to centralized control by a few major corporations. This concentration of power intensifies biases within AI models, resulting from inadequate governance and oversight mechanisms. Additionally, it limits public involvement and heightens concerns about the integrity of model generation. Such monopolistic control over data and AI outputs threatens both innovation and fair data usage, as users inadvertently contribute data that primarily benefits these corporations. In this work, we propose AIArena , a blockchain-based decentralized AI training platform designed to democratize AI development and alignment through on-chain incentive mechanisms. AIArena fosters an open and collaborative environment where participants can contribute models and computing resources. Its on-chain consensus mechanism ensures fair rewards for participants based on their contributions. We instantiate and implement AIArena on the public Base blockchain Sepolia testnet, and the evaluation results demonstrate the feasibility of AIArena in real-world applications. Our project page is available at https://train.flock.io/explore . Conference: Companion Proceedings of the ACM Web Conference 2025; April 28-May 2, 2025; Sydney, NSW, Australia Companion Proceedings of the ACM Web Conference 2025 (WWW Companion ’25), April 28-May 2, 2025, Sydney, NSW, Australia DOI: 10.1145/3701716.3715484 ISBN: 979-8-4007-1331-6/2025/04 CCS: Computing methodologies Artificial intelligence CCS: Computing methodologies Machine learning CCS: Theory of computation Design and analysis of algorithms † † cc-license: by 1. Introduction The growth of artificial intelligence (AI) has led to groundbreaking applications in many fields. However, it also brings challenges, especially because a few large companies control most of its development and use. This concentration of power increases bias in AI systems ( Challen et al., 2019 ) and reduces public involvement in important decisions. This lack of transparency can exacerbate issues such as misuse of AI technologies or unethical practices, as there is little external oversight in the AI centralized systems. Such centralized control also slows down innovation in AI ( Etzioni, 2016 ; Montes and Goertzel, 2019 ) and leads to the unfair use of user data, which mainly benefits these companies. These compounded challenges underline the urgent need for a shift to decentralized AI (DeAI) ( Wang et al., 2024 ) , where control and access are distributed across a broader and more inclusive range of participants. Blockchain’s decentralized nature can help address the above challenges by allowing multiple participants to collaborate on AI development without relying on a central authority ( Chen et al., [n. d.] ; Dong et al., 2024 ) . Smart contracts ( Zou et al., 2019 ) can automate processes such as distributing rewards or verifying contributions, fostering fairness and efficiency. Blockchain also enhances data integrity, as participants can verify the source and ownership of data, preventing unauthorized use and ensuring compliance with ethical standards ( Salah et al., 2019 ) . Figure 1. Overview of AIArena , a blockchain-based decentralized AI training system. In this work, we propose AIArena , a blockchain-based platform designed to decentralize AI training in an open and collaborative environment, where participants can contribute models and computing power. Specifically, AIArena leverages on-chain consensus mechanisms to ensure that only valid contributions are rewarded, while incentivization models encourage active and meaningful participation to prevent issues such as free-riding or fake identities. Contributions. Our contributions are summarized as follows: • We propose AIArena , a blockchain-based decentralized platform for AI training. AIArena enables the AI model to be optimized directly on users’ devices using their own data or publicly available datasets, and finally derives the optimal model based on blockchain consensus. By leveraging blockchain, AIArena ensures sustained machine learning (ML) contributor (i.e., training node and validator) engagement and provides fair rewards based on their contributions in enhancing models. • We introduce a delegation stake design for AIArena , enabling users with limited computational resources to participate in model generation by delegating their stake to preferred ML participants. This design encourages broader involvement in the ML model generation process, further promoting decentralization. • We instantiate and implement AIArena on the public Base blockchain Sepolia testnet 1 1 1 https://sepolia.basescan.org/address/0x7b6bde1d173eb288f390ff36e21801f42c4d8d91 . Over an approximate 7 -month period, AIArena has engaged 603 training nodes, 1{,}051 validators, and 63{,}265 delegators, collaboratively generating 18{,}656 models across 16 training tasks. These results highlight the practical feasibility of AIArena in real-world applications. • We evaluate three popular tasks performed by AIArena contributors. Their models not only outperform baseline models but also exceed the performance of larger state-of-the-art (SOTA) models. Notably, in the code co-creation task, our contributors collaboratively curated the largest existing Move code dataset, along with their corresponding code instructions and comments. 2. AIArena System Design 2.1. Participants Fig. 1 shows the overview of AIArena system. There are various categories of participants in the AIArena system: • Task Creators: Task creators are responsible for defining training tasks and their specific requirements, including selecting and designing appropriate learning algorithms for model training and validation. To promote decentralization, the created tasks can be reviewed and verified by the stakeholders within AIArena . • Training Nodes: Given a created task, training nodes are responsible for training the model on their service using public data. To participate, training nodes must stake blockchain assets or tokens to establish their eligibility. Training nodes will receive rewards based on their stake amount and performance. • Validators: Validators are responsible for assessing the work completed by training nodes and submitting validation scores that influence the distribution of rewards. To participate, validators must stake tokens, which not only allows them to validate assigned tasks but also ensures fair task allocation. • Delegators: Delegators contribute to the AIArena system by supporting the staking process of other participants without directly engaging in task training or validation. They stake tokens on behalf of other participants, boosting the delegatees’ capacity to earn greater rewards. In return, delegators share in the rewards earned by their delegatees, according to predefined algorithms that reflect their staked contributions. 2.2. Training and Validation 2.2.1. Training We consider the dataset held by the training node, \mathcal{D}_{\text{local}} , which contains locally sourced data samples, comprising feature set X and label set Y , with each sample x_{i}\in X corresponding to a label y_{i}\in Y . We define a predictive model f , aiming to learn patterns within \mathcal{D} such that f(x_{i})\approx y_{i} . To quantify the prediction metric, e.g., accuracy, the task trainer will introduce a loss function L(f(x_{i}),y_{i}) , assessing the discrepancy between predictions f(x_{i}) and actual labels y_{i} . A generic expression for this function is: L=\frac{1}{N}\sum_{i=1}^{N}l(f(x_{i}),y_{i}) , where N denotes the total sample count, and l signifies a problem-specific loss function, e.g., mean squared error or cross-entropy loss. The optimization goal is to adjust the model parameters \theta to minimize L , typically through algorithms such as gradient descent: \theta_{\text{new}}=\theta_{\text{old}}-\eta\nabla_{\theta}L , where \eta represents the learning rate, and \nabla_{\theta}L the gradient of L with respect to \theta . Utilizing the aggregated dataset \mathcal{D} , parameter \theta is iteratively updated to reduce L , consequently improving the model’s predictive accuracy. This optimization process is conducted over a predefined number of epochs E , each epoch consisting of

Read the complete text