Federated Learning for Privacy-Preserving Data Collaboration: A Threat-Aware Synthesis and Privacy Stack Selection Framework

Modern machine learning increasingly depends on data distributed across institutions and user devices, yet centralising sensitive records can create privacy, governance, and commercial barriers. Federated learning (FL) changes this architecture by retaining raw data at local clients and exchanging model updates instead. Data locality, however, is not equivalent to a complete privacy guarantee: gradients, updates, participation signals, aggregates, model parameters, and outputs may remain information-bearing artefacts. This paper presents a structured, literature-grounded synthesis of privacy threats and privacy-enhancing mechanisms in FL. A 50-work evidence set is organised through nine coding dimensions spanning deployment regime, data partitioning, adversary, exposed artefact, privacy objective, mechanism, trust assumption, operational overhead, and residual risk. Differential privacy, secure aggregation, homomorphic encryption and secure multi-party computation, trusted execution environments, and private entity-alignment mechanisms are compared as components of layered architectures rather than interchangeable solutions. From the synthesis, we define privacy objective–artefact misalignment: a design failure in which the chosen mechanism protects a different artefact or adversary surface from the deployment's stated privacy objective. We then propose a Threat-Aware Privacy Stack Selection Framework with an explicit Stack Incompleteness Test that checks coverage, adversary-assumption compatibility, mechanism interaction, and deployment feasibility before a candidate stack is accepted. The framework is a literature-derived analytical decision aid, not an optimality proof, quantitative risk calculator, or empirically validated automated selector. The central proposition is that data locality is an architectural property, whereas privacy is a threat-dependent system property.