Abstract
This paper addresses privacy protection in decentralized ArtificialIntelligence (AI) using Confidential Computing (CC) within the Atoma Network, adecentralized AI platform designed for the Web3 domain. Decentralized AIdistributes AI services among multiple entities without centralized oversight,fostering transparency and robustness. However, this structure introducessignificant privacy challenges, as sensitive assets such as proprietary modelsand personal data may be exposed to untrusted participants. Cryptography-basedprivacy protection techniques such as zero-knowledge machine learning (zkML)suffers prohibitive computational overhead. To address the limitation, wepropose leveraging Confidential Computing (CC). Confidential Computingleverages hardware-based Trusted Execution Environments (TEEs) to provideisolation for processing sensitive data, ensuring that both model parametersand user data remain secure, even in decentralized, potentially untrustedenvironments. While TEEs face a few limitations, we believe they can bridge theprivacy gap in decentralized AI. We explore how we can integrate TEEs intoAtoma's decentralized framework.