LaserTag: A Tool for Autonomous XAI-Guided Physical Adversarial Perturbations in Industrial Vision Pipelines
Year: 2026
Type: conference-paper
Abstract: We introduce LaserTag, an open-source toolchain for studying physical-world adversarial effects on industrial vision pipelines. The system couples lightweight classifiers with XAI methods and a reproducible pan–tilt laser actuation rig to identify salient image regions and probe them with controlled optical perturbations. Designed for dependability research, it supports interactive operation, automated logging, and repeatable experimentation. We evaluate LaserTag on three representative industrial controllers, highlighting practical vulnerabilities in vision-based status recognition. Rather than optimizing attacks, the toolchain provides an auditable testbed for autonomous XAI-guided targeting, surrogate-model transferability analysis, and countermeasure assessment. Experiments show that XAI-guided laser perturbations reliably induce both failure masking and spurious alarms, while several digital defenses—including color-channel suppression, highlight clipping, and laser inpainting—successfully restore correct predictions across both main and surrogate models. Code and models are released for reproducible research.
Authors: Gustavo Sánchez, Li Wei, Veit Hagenmeyer
Institution: Karlsruhe Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7167671611