Robust Few-Shot Learning Without Using Any Adversarial Samples.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning is crucial due to high data acquisition costs.
• Existing methods often neglect robustness against adversarial noise.
• Current adversarial few-shot methods are computationally intensive.

Purpose of the Study:

• To develop a computationally efficient and robust few-shot learning method.
• To improve adversarial accuracy without generating adversarial samples during training.
• To enhance model resilience against data perturbations.

Main Methods:

• Employs self-distillation for high-level feature matching between base class data and low-frequency samples during pretraining.
• Fine-tunes the model on novel classes, enhancing low-frequency query set feature discriminability via cosine similarity.
• Avoids the generation of adversarial samples in training episodes.

Main Results:

• Achieved substantial improvements in adversarial accuracy (60.55% on PGD, 62.05% on auto attack) on the CIFAR-FS dataset in a one-shot setting.
• Demonstrated a minor drop in clean accuracy compared to baseline methods.
• Achieved faster training times than state-of-the-art adversarial meta-learning methods.

Conclusions:

• The proposed method offers a simple, effective, and computationally efficient solution for robust few-shot learning.
• The technique enhances adversarial robustness without compromising clean performance significantly.
• This approach presents a practical alternative for real-world applications requiring resilient few-shot models.