Security Versus Accuracy: Trade-Off Data Modeling to Safe Fault Classification Systems.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Industrial Fault Classification

Background:

• Data-driven fault classification systems are widely used but vulnerable to adversarial attacks.
• Adversarial security is critical for safety-critical industrial applications, presenting a trade-off with accuracy.
• Existing methods often struggle to balance security and performance.

Purpose of the Study:

• To investigate the security-accuracy trade-off in fault classification models.
• To propose a novel approach using hyperparameter optimization (HPO) to enhance adversarial robustness.
• To develop an efficient multi-objective, multi-fidelity Bayesian optimization (BO) algorithm for HPO.

Main Methods:

• Developed a new multi-objective, multi-fidelity Bayesian optimization algorithm named MMTPE.
• Applied MMTPE for hyperparameter optimization of machine learning models in fault classification.
• Evaluated the proposed algorithm and optimized models on safety-critical industrial datasets.

Main Results:

• MMTPE demonstrated superior efficiency and performance compared to other advanced optimization algorithms.
• Optimized fault classification models achieved competitive adversarial robustness, comparable to specialized defensive methods.
• The study provided insights into intrinsic model security properties and hyperparameter-security correlations.

Conclusions:

• Hyperparameter optimization offers a new perspective for enhancing adversarial security in fault classification.
• The MMTPE algorithm effectively balances the trade-off between accuracy and adversarial robustness.
• Optimized models show significant improvements in security for safety-critical industrial applications.