Privacy-Preserving Computer Vision for Industry: Three Case Studies in Human-Centric Manufacturing

TL;DR

Balancing operational utility and worker privacy is a key barrier to industrial computer vision adoption. The authors present a task-centric privacy-preserving framework that learns a lightweight obfuscation transforming data at the edge while preserving task performance, validated across ergonomics, woodworking, and human-aware AGV navigation use cases. They provide quantitative privacy-utility analyses and qualitative deployment feedback, highlighting configurability, transparency, and third-party validation as critical for real-world adoption. The work demonstrates readiness for real-world deployment and offers cross-domain guidelines for responsible, human-centric AI in manufacturing.

Abstract

The adoption of AI-powered computer vision in industry is often constrained by the need to balance operational utility with worker privacy. Building on our previously proposed privacy-preserving framework, this paper presents its first comprehensive validation on real-world data collected directly by industrial partners in active production environments. We evaluate the framework across three representative use cases: woodworking production monitoring, human-aware AGV navigation, and multi-camera ergonomic risk assessment. The approach employs learned visual transformations that obscure sensitive or task-irrelevant information while retaining features essential for task performance. Through both quantitative evaluation of the privacy-utility trade-off and qualitative feedback from industrial partners, we assess the framework's effectiveness, deployment feasibility, and trust implications. Results demonstrate that task-specific obfuscation enables effective monitoring with reduced privacy risks, establishing the framework's readiness for real-world adoption and providing cross-domain recommendations for responsible, human-centric AI deployment in industry.

Figures (7)

• Figure 1: Conceptual overview of the proposed task-centric, privacy-preserving vision framework. Raw sensor data are processed within a secure edge environment through a configurable privacy mechanism that filters task-relevant visual features before on-device or cloud-based analysis. This design embeds data minimization into the sensing process, aligning with human-centric privacy principles central to Industry 5.0.
• Figure 2: Overview of the adversarial framework. The Obfuscator ($O$) transforms the input image to preserve utility for the task model ($U$) while limiting the Deobfuscator’s ($D$) ability to reconstruct the original frame.
• Figure 3: Example frames from the three industrial use cases. Left: (a) Ergonomics. Right: (b) Woodworking and (c) AGV pedestrian detection.
• Figure 4: Original and obfuscated with detection examples for each use case. Note that detections are made by the original models that have not been retrained on obfuscated data.
• Figure 5: Privacy-utility curves for the three use-cases. We compare our obfuscator against basic techniques such as whole-image blurring and a two-step technique where persons are first detected and then blurred. Our technique achieves a more optimal privacy-utility tradeoff than these basic techniques.