Interpolated Adversarial Training: Achieving robust neural networks without sacrificing too much accuracy

Area of Science:

• Deep Learning
• Machine Learning Security
• Computer Vision

Background:

• Adversarial robustness is crucial for deep learning systems.
• Standard adversarial training methods often degrade generalization performance on unperturbed data.
• This trade-off impacts the practical adoption of robust models.

Purpose of the Study:

• To introduce Interpolated Adversarial Training (IAT) as a novel method.
• To improve adversarial robustness while maintaining high generalization performance.
• To address the performance degradation issue in adversarial training.

Main Methods:

• Employing interpolation-based training techniques within the adversarial training framework.
• Applying the proposed method to the CIFAR-10 dataset.
• Conducting mathematical analysis to validate the approach.

Main Results:

• IAT achieved a standard test error of 6.45% on CIFAR-10, compared to 12.32% for standard adversarial training.
• The relative increase in standard error for the robust model was reduced from 178.1% to 45.5% using IAT.
• Maintained adversarial robustness while significantly improving generalization.

Conclusions:

• Interpolated Adversarial Training offers a superior balance between robustness and generalization.
• The method effectively mitigates the performance degradation associated with traditional adversarial training.
• IAT presents a promising direction for developing more practical and robust deep learning models.