Adversarial Robustness Enhancement for Deep Learning-Based Soft Sensors: An Adversarial Training Strategy Using...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Chemical Engineering

Background:

• Deep learning-based soft sensor (DLSS) models offer high prediction accuracy but are vulnerable to adversarial attacks.
• Existing adversarial training methods struggle with transfer gradient estimation and robust overfitting, limiting DLSS deployment.
• Adversarial attacks pose a significant risk to the safe and reliable application of DLSS in industrial processes.

Purpose of the Study:

• To propose a novel adversarial training approach, domain-adaptive adversarial training (DAAT), to enhance the adversarial robustness of DLSS.
• To address the limitations of existing adversarial training methods, specifically transfer gradient estimation and robust overfitting.
• To improve the practical applicability and safety of DLSS models in real-world manufacturing environments.

Main Methods:

• Developed a two-stage approach: historical gradient-based adversarial attack (HGAA) and domain-adaptive training.
• HGAA stabilizes gradient updates using historical information for stronger adversarial sample generation.
• Domain-adaptive training learns common features from adversarial and original samples to prevent overfitting and enhance robustness.

Main Results:

• The proposed DAAT method effectively enhances the adversarial robustness of DLSS models.
• DAAT achieves a balance between defending against adversarial samples and maintaining prediction accuracy on normal samples.
• A case study on silicon single-crystal growth demonstrated the practical effectiveness of DAAT.

Conclusions:

• DAAT offers an effective solution for improving the adversarial robustness of DLSS models.
• The method addresses key challenges in adversarial training, leading to more secure and reliable soft sensors.
• DAAT contributes to the safe deployment of advanced DLSS in critical industrial applications.