Working around rather than using it as intended: A think-aloud study of Differential Privacy library usability

Abstract

Differential Privacy (DP) guarantees privacy-preserving data analysis, yet adoption remains limited. We conducted a think-aloud study with 22 Python developers of varied experience—they attempted two standard Differential Privacy analytics tasks across IBM DiffPrivLib, PyDP, and OpenDP. We found developers resorted to undocumented workarounds—using low-level mechanisms directly, manually implementing functionality the library should provide, or relying on external tools for basic operations—as high-level APIs failed to address common use cases; suggesting a fundamental gap between how libraries are designed and how developers actually intend to use them. Our primary contribution is an empirically grounded taxonomy of workaround strategies: alternative procedures that developers improvise when intended library workflows do not fit their task. We trace these recurring workaround strategies each to specific, DP -relevant library design and documentation gaps. We suggest actionable recommendations for library maintainers seeking to make Differential Privacy usable without forcing developers into privacy-compromising—solution patterns.