Position: Verifiable Data Minimization is a Prerequisite for Responsible, Privacy-Preserving Industrial Vision
Abstract
The adoption of computer vision to drive industrial efficiency and safety creates a persistent tension between operational utility and worker privacy. Current privacy measures, such as post-hoc blurring, are fundamentally flawed: they depend on the error-prone detection of sensitive attributes yet do not prevent the data from being repurposed for unintended inferences. We posit that industrial computer vision must shift from ”hiding secrets” to verifiable data minimization. Formalized through the Information Bottleneck principle, we advocate for sensing pipelines that are optimized to capture only the features necessary for a declared task (e.g., pose estimation), without requiring labels for what is sensitive. This bounds the system’s capacity for unauthorized inferences such as re-identification or function creep. We outline an implementation path using modular edge processing and trusted execution environments, and identify key research challenges for making verifiable purpose limitation a prerequisite for responsible industrial AI.