Regularization Meets Enhanced Multi-Stage Fusion Features: Making CNN More Robust against White-Box Adversarial...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Regularization is a key technique in adversarial defense for Convolutional Neural Networks (CNNs).
• Existing methods lack clarity on which CNN features are optimal for regularization.

Purpose of the Study:

• To propose a novel network, the Enhanced Multi-Stage Feature Fusion Network (EMSF²Net), for improved adversarial defense.
• To investigate the impact of feature enhancement and fusion on adversarial robustness.

Main Methods:

• Developed EMSF²Net incorporating Multi-Stage Feature Enhancement (MSFE) and Multi-Stage Feature Fusion (MSF²).
• Implemented a regularization operation on both original and fused features during training.
• MSFE enhances features by channel-wise multiplication; MSF² fuses features across stages.

Main Results:

• EMSF²Net significantly improves adversarial robustness of CNNs by incorporating regularization on enhanced multi-stage features.
• Experimental results on the CIFAR-10 dataset demonstrate the method's effectiveness against white-box attacks.

Conclusions:

• The proposed EMSF²Net effectively enhances adversarial robustness in CNNs.
• Feature enhancement and fusion combined with regularization are crucial for defending against sophisticated attacks.