Universal attention guided adversarial defense using feature pyramid and non-local mechanisms

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Neural Networks (DNNs) are vulnerable to adversarial examples, limiting their use in secure applications.
• Existing defenses lack universality, being effective only against specific attack types.
• Analyzing adversarial examples via model attention shifts offers a new defense perspective.

Purpose of the Study:

• To develop universal defense modules against adversarial attacks on DNNs.
• To classify attacks based on their impact on model attention.
• To enhance adversarial robustness with minimal model intrusion.

Main Methods:

• Classified adversarial attacks into attention-shifting and attention-attenuation categories.
• Proposed Feature Pyramid-based Attention Space-guided (FPAS) module for attention-shifting attacks.
• Developed Attention-based Non-Local (ANL) module for attention-attenuation attacks.
• Integrated FPAS and ANL modules into a Wide-ResNet model within a boosting framework.

Main Results:

• The integrated FPAS and ANL modules demonstrated synergistic defense capabilities.
• Significant improvements in defense rates were observed, even against patch-embedded adversarial examples.
• The proposed universal defense strategy achieved average defense rate enhancements of 5.47% and 7.74% over the baseline.

Conclusions:

• The novel defense modules provide comprehensive protection against diverse adversarial attacks.
• This universal defense strategy offers a cost-effective solution compared to mainstream methods.
• The approach is adaptable for integration with existing defenses to further improve adversarial robustness.