Understanding Adversarial Robustness From Feature Maps of Convolutional Layers.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial robustness in neural networks is crucial for reliable AI systems.
• Robustness depends on model capacity and anti-perturbation ability.
• Feature maps in convolutional layers are key to understanding anti-perturbation ability.

Purpose of the Study:

• To investigate the relationship between convolutional feature map size and adversarial robustness.
• To identify how pooling operations affect this relationship.
• To propose simple modifications for enhancing neural network robustness.

Main Methods:

• Theoretical analysis of feature maps before average pooling and max pooling.
• Proposing modifications like input upsampling and stride adjustments in downsampling operators.
• Empirical verification on benchmark architectures (AlexNet, VGG, ResNet18, PreActResNet18).

Main Results:

• Larger feature maps before average pooling correlate with improved resistance to perturbations.
• This correlation does not hold for max pooling.
• Proposed modifications led to significant improvements in natural accuracy and adversarial robustness.

Conclusions:

• Feature map size before average pooling is a critical factor for adversarial robustness.
• Simple architectural modifications can enhance neural network resilience.
• Findings offer practical insights for designing more robust neural networks.