Boosting Adversarial Training With Mitigating Hard Sample Interference.

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning

Background:

• Adversarial training (AT) improves model robustness but often sacrifices accuracy.
• Jointly optimizing accuracy and robustness can lead to trade-offs, especially with hard samples near decision boundaries.

Purpose of the Study:

• To propose a novel method, Mitigating Hard Sample Interference (MHSI), to alleviate accuracy-robustness sacrifices in adversarial training.
• To reduce instability caused by hard samples during AT without compromising performance.

Main Methods:

• Introduced a weighted adaptive (WA) mechanism to strengthen learning from clean samples and reduce hard sample impact.
• Designed a dynamic calibration (DC) strategy, analyzing gradient norms and Hessian matrices, to mitigate hard sample damage to robustness.
• Implemented MHSI from a sample-intervention perspective to address accuracy-robustness trade-offs.

Main Results:

• MHSI significantly improves both accuracy and robustness, outperforming state-of-the-art methods under glass-box attacks.
• Demonstrated effectiveness on CIFAR-10, CIFAR-100, Tiny ImageNet, and SVHN datasets.
• Achieved up to a 6.22% robustness gain over the AT baseline under an $l_{\infty }$ attack.

Conclusions:

• MHSI effectively enhances adversarial training by improving accuracy and robustness simultaneously.
• The proposed WA and DC mechanisms successfully mitigate hard sample interference.
• MHSI offers a promising direction for developing more robust and accurate deep learning models.