Protecting User Privacy in Distributed Learning Using Secure MultiParty Computation and Differential Privacy: Mitigating AI-Powered Identity Reconstruction Risks

Abstract

The study is devoted to the development of methods for protecting user data confidentiality in distributed machine learning systems using protocols that ensure privacy (MPyC, PySecureCircuit, NssMPClib). The work is based on the Twitter dataset (Sent140 from LEAF), containing 1.6 million tweets with sentiment tags, and aims to analyze the trade-off between model accuracy and data protection. The theoretical foundations of distributed learning, including federated learning and decentralized schemes, were analyzed, and privacy threats were systematized: attacks on gradients, metadata, models, aggregation servers, and inter-client attacks. To ensure security, cryptographic methods such as multi-party computation and differential privacy are used, implemented through gradient clipping and noise addition during training.