Usage Privacy in Data-Driven Systems: Theory and Experiments with Machine Learning Programs

The research proposes a method for formalizing and enforcing usage privacy restrictions in data-driven systems, focusing on restricting the use of proxies (strong predictors) for protected information types. The method associates proxy use with intermediate computations in the program, requires that results be strongly correlated with protected information and may causally affect the output, uses program analysis to detect it and provide witnesses, and then a normative judgment oracle determines impropriety and repairs the model. Experiments on social datasets show that the algorithm can discover proxy uses that existing techniques struggle to detect, and removes them while maintaining acceptable classification performance.