Interpolated Joint Space Adversarial Training for Robust and Generalizable Defenses.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Adversarial training (AT) is a key defense against adversarial attacks but often reduces standard accuracy and generalization to novel threats.
• Existing methods for improving generalization, like on-manifold or neural perceptual threat models, have limitations regarding data requirements or algorithmic complexity.

Purpose of the Study:

• To introduce a novel Joint Space Threat Model (JSTM) that effectively utilizes manifold information.
• To develop new adversarial attacks and defenses under JSTM, specifically the Robust Mixup strategy.
• To enhance model performance in terms of standard accuracy, robustness, and generalization.

Main Methods:

• Proposed the Joint Space Threat Model (JSTM) leveraging Normalizing Flow to exploit manifold information.
• Developed the Robust Mixup strategy to maximize adversity in interpolated images, enhancing robustness and preventing overfitting.
• Implemented Interpolated Joint Space Adversarial Training (IJSAT) as a novel defense mechanism.

Main Results:

• IJSAT demonstrated superior performance in standard accuracy, robustness, and generalization across benchmark datasets (CIFAR-10/100, OM-ImageNet, CIFAR-10-C).
• The Robust Mixup strategy effectively improved adversarial robustness and prevented overfitting.
• IJSAT proved flexible, serving as a data augmentation technique and enhancing existing AT approaches.

Conclusions:

• The proposed JSTM and IJSAT offer a significant advancement in adversarial robustness and generalization.
• IJSAT provides a versatile solution that can be integrated with existing methods to boost performance.
• This work addresses key limitations of current adversarial training defenses, paving the way for more resilient AI models.