Design and evaluation of a privacy-preserving distributed machine learning system

Thesis Abstract

The rapid proliferation of data-driven applications across sectors such as healthcare, finance, and government has heightened the need for effective, scalable, and secure machine learning (ML) systems that respect individuals’ privacy. Traditional centralized ML approaches necessitate aggregating large volumes of sensitive data in a single location, which exposes organizations and users to substantial privacy risks and legal compliance challenges. This study aims to design, implement, and critically evaluate a privacy-preserving distributed machine learning system that balances model accuracy with rigorous privacy guarantees. Specific objectives include (1) developing an architecture that integrates federated learning with cryptographic techniques such as secure multi-party computation and differential privacy; (2) implementing the proposed model in a simulated environment reflecting real-world data characteristics; and (3) assessing the system’s performance against key metrics including privacy assurance, model utility, computational efficiency, and scalability. Adopting a mixed-methods research design, the study combines quantitative performance evaluation with qualitative assessments of privacy-preserving effectiveness. The population comprises simulated datasets encapsulating healthcare records, financial transactions, and social media data, aggregated from publicly available repositories. A stratified sampling technique determines a sample size of 2000 data points, ensuring adequate diversity for contextually relevant results. Data collection instruments include synthetic datasets generated via established data synthesis techniques, calibrated to mirror real-world distributions, and cryptographic protocol implementations coded in Python and Java. The evaluation methodology employs statistical analysis techniques such as regression analysis to examine the relationship between privacy parameters and model accuracy, along with ANOVA to compare the performance across various privacy-preserving configurations. Additionally, thematic analysis is applied to expert reviews of the system’s privacy guarantees and usability. The anticipated findings indicate that the integrated approach significantly enhances privacy security for distributed ML models without substantially degrading predictive performance. Specifically, it is expected that models employing differential privacy with an epsilon value of 0.5, combined with secure multi-party computation, will maintain over 85% accuracy relative to traditional centralized models while providing provable privacy guarantees. Furthermore, the system is projected to demonstrate scalable computational efficiency, with training times increasing linearly relative to data volume, thus supporting large-scale deployment potential. Sensitivity analyses are anticipated to reveal trade-offs between privacy levels and model utility, informing optimal configuration parameters for specific use cases. The study’s contribution to knowledge lies in presenting a comprehensive framework for implementing practical, scalable, and robust privacy-preserving distributed machine learning systems. By integrating cryptographic and differential privacy techniques within a federated learning architecture, the research advances the design of systems that can operate securely in real-world settings—addressing a critical gap identified in prior literature, which often emphasizes either theoretical models without empirical validation or limited scalability analysis. The research further provides empirical evidence and practical guidelines for deploying privacy-enhanced ML solutions, informing policymakers, practitioners, and researchers engaged in secure data analytics. In conclusion, the study demonstrates that well-integrated privacy-preserving techniques can be effectively operationalized within distributed ML architectures, maintaining high levels of data utility and system efficiency. Recommendations for future work emphasize exploring adaptive privacy mechanisms tailored to dynamic data environments, extending the approach to emerging AI domains such as federated reinforcement learning, and conducting field deployments in real-world organizational contexts. Overall, the research affirms that secure, privacy-respecting distributed machine learning systems are both technically feasible and essential for fostering trust in data-centric innovations, paving the way for broader adoption of privacy-conscious AI technologies.