Adapting Few-Shot Classification via In-Process Defense.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot learning (FSL) methods often rely on pre-trained models, creating vulnerabilities to backdoor attacks.
• Backdoor attacks can cause misclassification by exploiting hidden triggers, significantly degrading model performance.

Purpose of the Study:

• To develop adaptive defenses against backdoor attacks specifically for few-shot learning scenarios.
• To introduce a novel 'in-process defense' mechanism that operates during model forward propagation.

Main Methods:

• A specialized stochastic process tailored to task characteristics is employed to safeguard classification models.
• Latent stochastic processes approximate task distributions, deriving task-level representations from the support set.
• This task-level representation guides feature extraction, leading to backdoor trigger mismatching and forming a parameter defense strategy.

Main Results:

• The proposed in-process defense effectively counters backdoor attacks in few-shot learning.
• Benchmark tests on Meta-Dataset demonstrate the approach's robustness against attack-induced incorrect feature extraction.
• The method shows improved adaptation capabilities for few-shot classification tasks.

Conclusions:

• The developed adaptive defense strategy successfully mitigates backdoor attacks in few-shot learning.
• The in-process defense mechanism provides a robust and adaptable solution for securing FSL models.
• This research contributes to the development of more secure and reliable few-shot learning systems.