Robust long-tailed recognition with distribution-aware adversarial example generation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial attacks pose a significant threat to machine learning models.
• Achieving adversarial robustness under long-tailed data distributions is challenging.
• Existing adversarial training methods often overlook imbalances in adversarial example generation.

Purpose of the Study:

• To investigate the impact of long-tailed distributions on the adversarial example generation phase.
• To propose a novel method for balancing adversarial example generation across different classes.
• To enhance adversarial robustness for models trained on imbalanced datasets.

Main Methods:

• Proposed Distribution-Aware Adversarial Example Generation (DAG) method.
• Introduced Virtual Example Creator (VEC) to balance adversarial perturbations.
• Utilized Gradient-Guided Calibrator (GGC) to focus on tail classes based on generation quality.

Main Results:

• Demonstrated that adversarial example generation quality is inferior for tail classes.
• DAG effectively balances adversarial example generation across head and tail classes.
• Achieved superior performance compared to existing methods on long-tailed adversarial benchmarks.

Conclusions:

• The proposed DAG method effectively addresses the imbalance in adversarial example generation for long-tailed distributions.
• DAG significantly improves adversarial robustness, particularly for underrepresented classes.
• DAG represents a substantial advancement in robust machine learning for imbalanced datasets.