Unveiling ML Developers' Perceived and Actual Barriers of Privacy-Preserving Machine Learning Techniques

Differential privacy (DP) has become a common approach to privacy-preserving machine learning, but putting its guarantees into practice requires developers to translate privacy goals into concrete implementation choices. We present an exploratory study with 10 ML developers combining structured interviews with a hands-on implementation task using the Opacus library to directly compare anticipated and actual barriers to DP implementation. Before the task, participants expected challenges around privacy-utility tradeoffs and documentation quality. During implementation, participants instead struggled with API-level friction and identifying which parameters controlled privacy. Most strikingly, 6 of 10 participants set a target epsilon value expressing their intended privacy goal, then used Opacus' make_private API in a way that did not use epsilon to determine the resulting privacy guarantee, silently decoupling their stated privacy intent from their actual training configuration. While participants generally recognized that more sensitive data warranted stronger protection, this awareness rarely translated into principled parameter choices. Our findings point to a gap between developers' privacy intentions and their ability to operationalize them, suggesting that PPML tools need clearer interpretive guidance and more explicit feedback linking stated privacy intentions to actual behavior.