SoK: Towards Privacy-Centric Collaborative Machine Learning—A Classification Framework for Privacy Solutions

Le résumé fourni par la source

Abstract Machine learning (ML) relies on data to train models and make predictions, but when multiple parties collaborate, data privacy concerns arise. Organisations often resort to data silos to protect sensitive information and intellectual property, limiting the potential benefits of joint learning. Collaborative Machine Learning (CML) seeks to address this challenge by enabling parties to share insights rather than raw data, enhancing privacy while maintaining model performance. However, research has shown that CML architectures remain vulnerable to attacks such as data reconstruction and inference, primarily due to shared elements like metadata, model parameters, and decisions, necessary for insight exchange. In response, various privacy-enhancing techniques have been proposed, each offering privacy protection for individual shared elements instead of all. In this Systematisation of Knowledge (SoK) paper, we analyse and categorise existing CML privacy solutions to assess their objectives. Based on this, we introduce a framework to systematically classify privacy guarantees across different approaches, providing a structured taxonomy. By defining distinct privacy classes and mapping solutions accordingly, our work seeks to empower stakeholders to make informed decisions when adopting CML strategies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.