Social metaverse is a shared digital space combining a series of interconnected virtual worlds for users to play, shop, work, and socialize. In parallel with the advances of artificial intelligence (AI) and growing awareness of data privacy concerns, federated learning (FL) is promoted as a paradigm shift towards privacy-preserving AI-empowered social metaverse. However, challenges including privacy-utility tradeoff, learning reliability, and AI model thefts hinder the deployment of FL in real metaverse applications. In this paper, we exploit the pervasive social ties among users/avatars to advance a social-aware hierarchical FL framework, i.e., SocialFL for a better privacy-utility tradeoff in the social metaverse. Then, an aggregator-free robust FL mechanism based on blockchain is devised with a new block structure and an improved consensus protocol featured with on/off-chain collaboration. Furthermore, based on smart contracts and digital watermarks, an automatic federated AI (FedAI) model ownership provenance mechanism is designed to prevent AI model thefts and collusive avatars in social metaverse. Experimental findings validate the feasibility and effectiveness of proposed framework. Finally, we envision promising future research directions in this emerging area.

翻译：社交元宇宙是一个结合了系列互联虚拟世界的共享数字空间，供用户娱乐、购物、工作和社交。随着人工智能（AI）的进步以及数据隐私问题意识的增强，联邦学习（FL）被推广为向隐私保护型AI赋能社交元宇宙的范式转变。然而，隐私-效用权衡、学习可靠性及AI模型窃取等挑战阻碍了FL在实际元宇宙应用中的部署。本文利用用户/虚拟化身间的普遍社交关系，提出了一种社交感知的分层联邦学习框架SocialFL，以实现社交元宇宙中更好的隐私-效用权衡。随后，基于区块链设计了一种无需聚合器的鲁棒联邦学习机制，该机制采用新型区块结构及改进的共识协议，具备链上链下协同特性。此外，基于智能合约与数字水印，设计了一种自动化联邦AI（FedAI）模型所有权追溯机制，以防止社交元宇宙中的AI模型窃取及合谋虚拟化身行为。实验验证了所提框架的可行性与有效性。最后，我们展望了这一新兴领域中有前景的未来研究方向。