Mitigating low-frequency bias: Feature recalibration and frequency attention regularization for adversarial...

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks.
• Adversarial training (AT) improves robustness but introduces a low-frequency feature bias.
• This bias neglects crucial high-frequency details, impacting model performance.

Purpose of the Study:

• To address the low-frequency bias in adversarial training.
• To enhance the adversarial robustness of DNNs.
• To improve the capture of high-frequency features for better semantic understanding.

Main Methods:

• Propose High-Frequency Feature Disentanglement and Recalibration (HFDR) module.
• Implement frequency attention regularization to harmonize feature extraction.
• Separate and recalibrate frequency-specific features to capture latent semantic cues.

Main Results:

• HFDR consistently enhances adversarial robustness across datasets (CIFAR-10, CIFAR-100, ImageNet-1K).
• Achieved 2.89% gain on CIFAR-100 (WRN34-10) and 3.09% on ImageNet-1K.
• Demonstrated 4.89% gain on ViT-B against AutoAttack, showing adaptability to CNNs and Transformers.

Conclusions:

• HFDR effectively mitigates the low-frequency bias in adversarial training.
• The proposed method significantly improves DNN robustness against adversarial attacks.
• HFDR is adaptable to various architectures, including convolutional and transformer models.