Appendix A. Considerations for Sensitive Data within Machine Learning Datasets
The content of this appendix, written by the author and Brad Svee, was published as a solution paper on the Google Cloud Platform documentation website.
When you are developing a machine learning (ML) program, it’s important to balance data access within your company against the security implications of that access. You want insights contained in the raw dataset to guide ML training even as access to sensitive data is limited. To achieve both goals, it’s useful to train ML systems on a subset of the raw data, or on the entire dataset after partial application of any number of aggregation or obfuscation techniques.
For example, you might want your data engineers to train an ML model to weigh customer feedback on a product, but you don’t want them to know who submitted the feedback. However, information such as delivery address and purchase history is critically important for training the ML model. After the data is provided to the data engineers, they will need to query it for data exploration purposes, so it is important to protect your sensitive data fields before making it available. This type of dilemma is also common in ML models that involve recommendation engines. To create a model that returns user-specific results, you typically need access to user-specific data.
Fortunately, there are techniques you can use to remove some sensitive data from your datasets while still training effective ...