Fig 1.
Illustration of human-AI perception discrepancy in alpha channel attacks.
Fig 2.
Comparative evaluation of adversarial attack methods across four key criteria.
Table 1.
Model architectures, training parameters, and hardware settings.
Fig 3.
Visual deception of alpha channel attack in industrial inspection.
Fig 4.
Output deviation of Fast-GAN under alpha channel attack.
Fig 5.
Training flowchart of YOLO detection model under poisoned data.
Fig 6.
Detection failure of YOLO model induced by alpha channel attack.
Fig 7.
Output deviation of DeepSeek-VL on alpha attacked image.
Fig 8.
Misinterpretation of alpha attacked image by ChatGPT-4.
Fig 9.
Misunderstanding of alpha attacked image by Kimi multimodal model.
Table 2.
Detection performance (mAP@50/ mAP@50–95) of each defect category under different attack methods.
Fig 10.
Impact of alpha-channel attack on GAN training (measured by FID score).
Table 3.
Language consistency evaluation under alpha-channel attacks.
Table 4.
Multimodal alignment degradation measured by CLIP score.
Table 5.
Semantic drift in text descriptions caused by alpha-channel attacks.
Table 6.
Comparative impact of alpha channel, FGSM, and boundary attacks across diverse models and tasks.
Fig 11.
Flow chart of alpha channel attack image detection.
Fig 12.
Detection pipeline and quantitative metrics for alpha channel attack.
Fig 13.
ROC curve for alpha channel attack detection.
Table 7.
Performance metrics of the alpha-channel-based detection model.
Table 8.
Performance comparison of main adversarial defense methods across detection quality, efficiency, and robustness dimensions.
Fig 14.
Adversarially perturbed weld defect dataset.
Fig 15.
Grayscale histogram of weld data under attack.
Fig 16.
MSE heatmap of weld data under attack.
Fig 17.
Adversarially perturbed PCB defect dataset.
Fig 18.
Grayscale histogram of PCB data under attack.
Fig 19.
MSE heatmap of PCB data under attack.
Fig 20.
Adversarially perturbed steel surface defect dataset.
Fig 21.
Grayscale histogram of steel surface data under attack.
Fig 22.
MSE heatmap of steel surface data under attack.