Stylized Adversarial Defense.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Convolutional Neural Networks (CNNs) are vulnerable to subtle input image perturbations, leading to misclassifications.
• Existing adversarial training methods often rely solely on class-boundary information, limiting their effectiveness.

Purpose of the Study:

• To develop a more robust adversarial training method for CNNs.
• To enhance model resilience against imperceptible image changes and data distributional shifts.

Main Methods:

• Proposing an adversarial training approach that leverages feature space information (style, content, class-boundary) to craft stronger adversaries.
• Employing a multi-task objective with deep supervision for multi-scale feature extraction.
• Implementing a max-margin objective to differentiate between source images, adversaries, and target images.

Main Results:

• The proposed method demonstrates superior robustness against adversarial attacks compared to state-of-the-art defenses.
• The robust model generalizes well to naturally occurring corruptions and data distributional shifts.
• The approach maintains high accuracy on clean, unperturbed data.

Conclusions:

• Exploiting feature space information in adversarial training significantly improves CNN robustness.
• The proposed method offers a promising direction for building more resilient deep learning models.
• This technique enhances generalization capabilities beyond standard adversarial robustness.