Alleviating noise memorization for adversarially robust few-shot learning.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Few-Shot Learning (FSL) enables models to learn new classes from limited data.
• Existing FSL models are vulnerable to adversarial attacks, especially with scarce data.
• Adversarial training, a common defense, can cause models to memorize noise, hindering generalization.

Purpose of the Study:

• To address the overlooked vulnerability of FSL models to adversarial attacks.
• To propose a novel method that mitigates noise memorization in FSL.
• To enhance the generalization capabilities of FSL models under adversarial conditions.

Main Methods:

• Introducing Alleviation of Noise Memorization (ANM), a novel approach for robust FSL.
• Implementing Adaptive Label Smoothing for more flexible supervision.
• Incorporating Robust Weight Learning to improve model stability against adversarial noise.

Main Results:

• ANM effectively reduces the memorization of adversarial noise.
• The proposed method significantly improves the generalization ability of FSL models.
• Experimental results demonstrate ANM's superior performance compared to current benchmarks.

Conclusions:

• Adversarial training with hard labels can negatively impact FSL model generalization by promoting noise memorization.
• ANM offers a promising solution to enhance FSL robustness and generalization.
• The findings highlight the importance of addressing noise memorization for secure and reliable FSL applications.