Defensive Few-Shot Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot learning models are vulnerable to adversarial attacks.
• Existing adversarial defense methods are ineffective in few-shot settings due to violated distribution assumptions.
• A new framework is needed to address adversarial robustness in few-shot learning.

Purpose of the Study:

• To introduce and address the problem of defensive few-shot learning (DFSL).
• To develop a general framework for robust few-shot model training against adversarial attacks.
• To investigate methods for transferring adversarial defense knowledge and reducing distribution gaps in few-shot scenarios.

Main Methods:

• Proposed an episode-based adversarial training mechanism for knowledge transfer across sample distributions.
• Introduced feature-wise and prediction-wise distribution consistency criteria within few-shot tasks.
• Developed a general Defensive Few-Shot Learning (DFSL) framework.

Main Results:

• The proposed DFSL framework effectively enhances the robustness of few-shot models against adversarial attacks.
• Experimental results validate the framework's ability to transfer adversarial defense knowledge.
• The methods successfully narrow the distribution gap between clean and adversarial examples in few-shot tasks.

Conclusions:

• The developed DFSL framework provides an effective solution for building robust few-shot models against adversarial threats.
• The proposed training mechanism and distribution consistency criteria are key to achieving adversarial robustness in few-shot learning.
• This work opens new avenues for research in secure and reliable few-shot learning applications.