KAT: A Knowledge Adversarial Training Method for Zero-Order Takagi-Sugeno-Kang Fuzzy Classifiers.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Fuzzy Systems

Background:

• Adversarial training enhances classifier generalization but can be sensitive to adversarial samples.
• Existing methods often rely on input or output perturbations.

Purpose of the Study:

• To develop a novel knowledge adversarial attack model for zero-order Takagi-Sugeno-Kang (TSK) fuzzy classifiers.
• To improve generalization capability and interpretability while avoiding sensitivity to inappropriate adversarial samples.

Main Methods:

• Proposed a knowledge adversarial attack model perturbing interpretable fuzzy rules.
• Mimicked human-like thinking by considering knowledge-oblivion and/or knowledge-bias in fuzzy rules.
• Devised a knowledge adversarial training (KAT) method with dynamic regularization.

Main Results:

• KAT demonstrated promising generalization performance, interpretability, and fast training.
• Effectiveness validated on 15 benchmarking datasets from UCI and KEEL.
• The method avoids direct use of adversarial samples, unlike perturbation-based methods.

Conclusions:

• The novel KAT method offers a robust approach to training interpretable fuzzy classifiers.
• KAT effectively enhances generalization without the pitfalls of traditional adversarial sample generation.
• The approach is theoretically justified for strong generalization capabilities.