Domain Knowledge Alleviates Adversarial Attacks in Multi-Label Classifiers.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Adversarial attacks and defenses are well-studied for single-label classification.
• Multi-label classification presents unique challenges for detecting adversarial examples.
• Domain knowledge offers potential for identifying incoherent predictions.

Purpose of the Study:

• To investigate the use of domain knowledge constraints for detecting adversarial examples in multi-label classification.
• To develop a semi-supervised learning framework incorporating first-order logic knowledge as constraints.
• To evaluate the effectiveness of domain knowledge in enhancing classifier robustness against adversarial attacks.

Main Methods:

• Converting domain knowledge into first-order logic constraints.
• Integrating these constraints into a semi-supervised learning problem.
• Developing a constrained classifier that enforces domain knowledge.
• Evaluating the classifier's ability to reject incoherent predictions, including adversarial examples.

Main Results:

• Domain-knowledge constraints effectively detect adversarial examples, even without prior knowledge of attacks.
• The proposed method naturally rejects samples with incoherent predictions.
• Experimental analysis shows constraints improve adversarial example detection, particularly when unknown to attackers.
• An adaptive attack exploiting constraint knowledge was implemented and compared to state-of-the-art attacks.

Conclusions:

• Incorporating domain knowledge as constraints is a promising approach for robust multi-label classification.
• This method offers a natural way to identify and reject adversarial examples by enforcing learned relationships.
• The findings suggest a significant step towards building more resilient multi-label classifiers against sophisticated attacks.