Journal of Data Science

Privacy-preserving machine learning methods seek to train useful models that do not disclose information about the data on which they were trained. Such methods are vital when organizations train neural networks on sensitive individual-level data and seek to release the models publicly. Their goal poses a trade-off between predictive performance (utility) and privacy protection. That trade-off makes privacy-preserving machine learning methods difficult to apply in practice, usually requiring extensive iteration and hyperparameter tuning. Yet, practitioners often have little guidance for navigating competing statistical, computational, and privacy demands. We present an implementation algorithm for the Stochastic Weight Averaging–Gaussian Pseudo Posterior Mechanism (SWAG-PPM), a Bayesian differentially private deep learning method. The implementation algorithm focuses on the joint tuning of two key hyperparameters whose interaction governs model convergence and the privacy–utility trade-off. We introduce novel diagnostic tools to evaluate convergence and guide hyperparameter adjustments. Using a transformer model for occupational injury classification, we demonstrate that diagnostic-guided tuning with SWAG-PPM can achieve strong privacy protection and utility. While our case study uses a specific dataset and model architecture, all methodological steps can apply to other settings where privacy risk is heterogeneously distributed.