Evidence-Based Multi-Feature Fusion for Adversarial Robustness.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep Neural Networks (DNNs) are vulnerable to adversarial attacks due to feature space perturbations.
• Existing defenses focus on denoising or recalibrating features but under-explore feature trustworthiness.
• Reliance on single, potentially contaminated features hinders DNN robustness.

Purpose of the Study:

• To propose a novel method for quantifying feature trustworthiness in DNNs.
• To enhance adversarial robustness by fusing reliable features from multiple network blocks.
• To improve DNN resilience against both white-box and black-box adversarial attacks.

Main Methods:

• Introduction of Evidence-based Multi-Feature Fusion (EMFF) for adversarial robustness.
• Utilizing evidential deep learning to quantify belief mass and uncertainty in features.
• Employing Dempster's rule for evidential fusion of trusted features across network blocks.

Main Results:

• EMFF demonstrates significant advantages over existing defense techniques in adversarial robustness.
• The method proves effective against both white-box and black-box attack scenarios.
• Integration with adversarial training strategies improves robustness across various architectures, including CNNs and Vision Transformers.

Conclusions:

• EMFF enhances DNN adversarial robustness by reliably fusing trustworthy features.
• The approach offers a cost-effective solution with minimal parameter increase.
• Evidential deep learning provides a robust framework for assessing feature reliability in adversarial contexts.