# Visual security defense for industrial inspection based on computer vision

**Authors:** Zhihao Jiang, Haotian Yuan, Chenrui Zeng, Liu Fu

PMC · DOI: 10.1371/journal.pone.0338835 · 2026-02-04

## TL;DR

This paper introduces a new security threat to industrial computer vision systems using hidden image layers and proposes a defense to detect it.

## Contribution

The paper introduces the Alpha Channel Attack and a lightweight defense mechanism with high detection accuracy.

## Key findings

- Alpha-channel perturbations significantly degrade detection and generation metrics without altering visible content.
- The proposed defense achieves an AUC of 0.998 in identifying adversarial samples.
- The attack affects various model architectures, including vision-language models.

## Abstract

As intelligent manufacturing advances, computer vision-based defect detection systems have become essential components of industrial automation. However, this progress has also revealed new security vulnerabilities. In this work, we identify and examine a stealthy adversarial vector—the Alpha Channel Attack—which exploits the often-ignored transparency layer in RGBA images to inject imperceptible perturbations, thereby evading both human perception and conventional preprocessing defenses.We evaluate this threat across diverse model architectures, including YOLOv5, FastGAN, and state-of-the-art vision-language models such as DeepSeek-VL2, ChatGPT-4o, and KIMI. Experimental results show that alpha-channel perturbations cause substantial degradation in detection, generation, and multimodal alignment metrics—including mAP, FID, BLEU, METEOR, and CLIP Score—while leaving the visible image content unchanged.To mitigate this invisible yet high-impact risk, we propose a lightweight detection mechanism that integrates histogram overlap and MSE analysis within the alpha channel. The framework achieves an AUC of 0.998, demonstrating strong capability in identifying adversarial samples under real-world constraints.Overall, this study reveals a critical blind spot in modern visual data pipelines and introduces both a novel threat model and an effective defense strategy, contributing to the development of more resilient industrial AI systems.

## Full-text entities

- **Diseases:** weld defect (MESH:D000013), GAN (MESH:D004829), channel (MESH:C538353), poisoning (MESH:D011041), YOLO (MESH:D054331), hallucination (MESH:D006212), screw defect (MESH:D012610), LLMs (MESH:D007806)
- **Chemicals:** steel (MESH:D013232), PCB (MESH:D011078), metal (MESH:D008670), GAN (-)
- **Species:** Homo sapiens (human, species) [taxon 9606], Felis catus (cat, species) [taxon 9685], Solanum lycopersicum (tomato, species) [taxon 4081], Canis lupus familiaris (dog, subspecies) [taxon 9615]

## Figures

50 figures with captions in the complete paper: https://tomesphere.com/paper/PMC12872028/full.md

---
Source: https://tomesphere.com/paper/PMC12872028