Methodological Framework for the Architecture of Data Anonymization for Machine Learning Tasks

Abstract

This paper presents a methodological framework for the architecture of a tabular-data anonymization system embedded into the lifecycle of corporate machine learning projects and data preparation workflows. We propose a process- and stage-based approach to designing an anonymization pipeline that establishes a unified terminology, requirements, and constraints, and formalizes rule profiles for pseudonymization, generalization, masking, and suppression across different attribute classes: direct identifiers, quasi-identifiers, and sensitive attributes. Building on the k-anonymity, l-diversity, and t-closeness models, we introduce "privacy checkpoints" at which attainment of target metric values, suppression rates, and the level of generalization are evaluated. At each checkpoint, a privacy report is generated containing the observed k, l, and t values, warnings, and explanatory notes, enabling an informed decision on whether a dataset can be admitted into the ML pipeline. The paper also shows how to pre-validate profiles and parameters on representative anonymized samples without accessing actual production datasets, thereby reducing disclosure risks at early approval stages. The framework further specifies roles and responsibility boundaries (data owner, data engineer, analyst/data scientist, ML engineer, information security specialist, and system administrator) and a three-tier system architecture with a web interface and an API suitable for integration with pipeline orchestrators. Treating rule profiles as versioned artifacts-alongside dataset versions, run parameters, metadata storage, operation logging, and periodic auditing-ensures reproducibility of training data preparation and end-to-end traceability of anonymization impacts on model quality. The framework can serve as a reference model for an initial pilot implementation and subsequent expansion to other data classes and privacy governance practices in ML projects .