Generalizable and Discriminative Representations for Adversarially Robust Few-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot image classification (FSIC) aims to build recognition systems with limited data, crucial for real-world applications.
• Existing deep learning models are often vulnerable to adversarial examples, even with extensive training data.
• Current adversarial FSIC research predominantly relies on meta-learning, which can be computationally intensive.

Purpose of the Study:

• To develop a robust and effective baseline for few-shot image classification against adversarial examples.
• To propose a method that learns discriminative representations without requiring tedious meta-task sampling.
• To generalize the approach to unseen adversarial FSIC tasks.

Main Methods:

• Introduced an adversarial-aware (AA) mechanism for supplementary supervision using feature-level distinctions.
• Designed an adversarial reweighting training strategy to address class imbalance with adversarial examples.
• Proposed a cyclic feature purifier in postprocessing to enhance robustness against unforeseen adversarial attacks.

Main Results:

• Achieved state-of-the-art robustness and natural performance in adversarial FSIC.
• Demonstrated superior transferability of feature embeddings, even against cross-domain adversarial examples.
• Outperformed existing methods on three standard benchmarks by a significant margin.

Conclusions:

• The proposed method offers a straightforward yet effective solution for adversarial-robust FSIC.
• The approach learns robust and discriminative features, enhancing generalization to new adversarial tasks.
• This work advances the field by providing a highly robust and performant FSIC system.