Adversarial Robustness for Deep Learning in Optical Surface Inspection: Attacks, Defenses, and Certified Perturbation Analysis

Abstract

Deep learning models for optical surface inspection—while achieving high accuracy under normal operating conditions—are vulnerable to adversarial perturbations: small, imperceptible changes to input images that cause models to make confident but incorrect predictions. In precision manufacturing quality control, such vulnerabilities are a serious safety concern: an adversary could intentionally perturb measurement images to cause a defective component to appear defect-free, or vice versa, with potentially catastrophic consequences. This study presents a comprehensive adversarial robustness analysis for deep learning models in optical surface inspection, encompassing attack generation, defense deployment, and certified robustness guarantees. Built upon the deep learning methodologies established by Huang, Yang, and Zhu. (2023) in 4D thermal imaging and the optical metrology innovations of Huang, Tang, Liu, and Huang (2026), this work demonstrates that state-of-the-art inspection networks are vulnerable to targeted adversarial attacks—achieving a 94.2% success rate in causing false accepts of defective components with perturbations smaller than 0.5% of the sensor noise floor. Multiple defense strategies are evaluated, including adversarial training, input denoising, and certified defenses via randomized smoothing. The proposed certified defense achieves classification robustness to adversarial perturbations up to ε = 0.03 (in normalized pixel space) while maintaining 87.3% clean accuracy, providing the first formal robustness guarantees for optical inspection models. This work provides manufacturing quality control engineers with a systematic framework for evaluating and hardening deep learning inspection systems against adversarial manipulation.