Multi-Label Adversarial Attack With New Measures and Self-Paced Constraint Weighting.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Adversarial attacks in multi-label learning involve complex optimization problems.
• Existing methods struggle with coarse measures for attack failure degree (AFD) and attack cost (AC), especially with conflicting constraints.

Purpose of the Study:

• To develop refined measures for AFD and AC in top-k adversarial attacks.
• To formulate novel optimization problems addressing constraint violations.
• To propose a self-paced weighting strategy for improved attack performance and stability.

Main Methods:

• Developed a Jaccard index-based measure for AFD and AC.
• Formulated optimization problems minimizing constraint violation using new AFD/AC measures.
• Implemented a self-paced weighting strategy for constraint prioritization.

Main Results:

• The Jaccard index measure better distinguishes attack failure degrees and costs.
• Weighting slack variables theoretically improves optimization outcomes.
• The self-paced weighting strategy yields larger attack gains and avoids optimization fluctuations.

Conclusions:

• The proposed refined measures and self-paced weighting strategy enhance adversarial attacks in multi-label learning.
• The method demonstrates superior performance and stability across benchmark datasets.