MeshDP-FSL: A Decentralized Model for Privacy-Preserving Collaborative Machine Learning and Data Governance

The surge of privacy regulations and the rapid expansion of distributed datasets have made conventional centralized learning models insufficient for contemporary applications. Conventional centralized learning algorithms encounter issues including privacy vulnerabilities, elevated communication expenses, and scalability constraints, rendering them inappropriate for distributed urban data environments. This paper introduces MeshDP-FSL, a decentralized technique that integrates the concepts of Data Mesh, Differential Privacy (DP), Local Differential Privacy (LDP), and Federated Split Learning (FSL). MeshDP-FSL aims to optimize the trade-offs between privacy and utility, minimize communication overhead, and improve scalability in collaborative machine learning. MeshDP-FSL ensures privacy on many levels by implementing Local Differential Privacy (LDP) at the data source, compressing activations prior to transmission, and integrating Differential Privacy (DP) noise into gradient updates at the server. The system utilizes Data Mesh concepts to facilitate decentralized data ownership while ensuring optimal privacy-utility trade-offs. A thorough mathematical examination substantiates the privacy assurances, adaptive noise calibration, and communication efficiency of MeshDP-FSL.