The ISA Global Cybersecurity Alliance recently published a guide on defending against adversarial AI attacks targeting machine vision systems used in industrial settings. These attacks involve subtle pixel perturbations or physical patches—such as stickers or projections—that are nearly imperceptible to humans yet can cause AI-driven inspection systems to misclassify parts, drop objects, or halt production. Even after retraining, fault-diagnosis networks remain vulnerable, potentially disrupting manufacturing operations across assembly lines, robotics, and visual quality control.

To counter these threats, the guide recommends a multi-layered defense strategy, including threat modeling, adversarial training (where models are trained with both clean and adversarially perturbed inputs), ensemble methods, runtime monitoring for anomalous inputs, and proactive simulation of attacks during development. By hardening models and continuously testing them against adversarial techniques, organizations can significantly boost resilience in safety-critical AI deployments.

‍

Source: https://gca.isa.org/blog/defending-against-adversarial-ai-attacks-on-machine-vision-systems